Search

GoogLeNet is a pre-trained Convolutional Neural Network (CNN) that allows transfer learning and has achieved high recognition rates in image classification tasks. A Recurrence Plot (RP) is an imaging method that depicts the recurrence of the state space system using coloured points and lines in 2D images. This work contributes to facilitating time series feature extraction by proposing a method that applies the GoogLeNet to time series images obtained with RP. The developed method is tested using simulated time series and selected time series from the M3 competition dataset. The results shows that the transfer learning approach allowed the extraction of business time series features by means of a GoogLeNet fine-tuned using 100 simulated time series. The combination of GoogLeNet and RPs outperforms the alternative and easier combination of GoogLeNet and plots of the time series and support the convenience of the RP transformation step. This application of deep learning techniques to business time series imaging offers opportunity for further developments. [ABSTRACT FROM AUTHOR]

Modern manufacturing is the world's largest and most automated industrial sector. The rise of Industry 4.0 technologies such as Big Data, Internet of Things (IoT) devices, and Machine Learning has enabled a better connection with machines and factory systems. Data harvesting allowed for a more seamless and comprehensive implementation of the knowledge-based decision-making process. New models that provide a competitive edge must be created by combining the Lean paradigm with the new technologies of Industry 4.0. This paper presents novel computer-based vision models for automated detection and classification of damaged packages from intact packages. In high-volume production environments, the package manual inspection process through the human eye consumes inordinate amounts of time poring over physical packages. Our proposed three different computer-based vision approaches detect damaged packages to prevent them from moving to shipping operations that would otherwise incur waste in the form of wasted operating hours, wasted resources and lost customer satisfaction. The proposed approaches were carried out on a data set consisting of package images and achieved high precision, accuracy, and recall values during the training and validation stage, with the resultant trained YOLO v7 model. [ABSTRACT FROM AUTHOR]

Tidal features are a key observable prediction of the hierarchical model of galaxy formation and contain a wealth of information about the properties and history of a galaxy. Modern wide-field surveys such as LSST and Euclid will revolutionize the study of tidal features. However, the volume of data will prohibit visual inspection to identify features, thereby motivating a need to develop automated detection methods. This paper presents a visual classification of ∼2000 galaxies from the DECaLS survey into different tidal feature categories: arms, streams, shells , and diffuse. We trained a convolutional neural network (CNN) to reproduce the assigned visual classifications using these labels. Evaluated on a testing set where galaxies with tidal features were outnumbered |$\sim 1:10$|⁠ , our network performed very well and retrieved a median |$98.7\pm 0.3$|⁠ , |$99.1\pm 0.5$|⁠ , |$97.0\pm 0.8$|⁠ , and |$99.4^{+0.2}_{-0.6}$|  per  cent of the actual instances of arm, stream, shell , and diffuse features respectively for just 20 per cent contamination. A modified version that identified galaxies with any feature against those without achieved scores of |$0.981^{+0.001}_{-0.003}$|⁠ , |$0.834^{+0.014}_{-0.026}$|⁠ , |$0.974^{+0.008}_{-0.004}$|⁠ , and |$0.900^{+0.073}_{-0.015}$| for the accuracy, precision, recall, and F1 metrics, respectively. We used a gradient-weighted class activation mapping analysis to highlight important regions on images for a given classification to verify the network was classifying the galaxies correctly. This is the first demonstration of using CNNs to classify tidal features into sub-categories, and it will pave the way for the identification of different categories of tidal features in the vast samples of galaxies that forthcoming wide-field surveys will deliver. [ABSTRACT FROM AUTHOR]

Data Augmentation (DA) is significant for Hyperspectral Image (HSI) classification especially in the case of limited labeled training data. Various DA models have been introduced to augment HSI data by using image processing techniques, prior knowledge or contextual information of HSI. However, spatial information in HSI could be inefficiently adopted in these DA models, leading to the lack of diversity and richness of the augmented data. To handle the issue, inspired by the recently proposed Principal Component Analysis based DA (PCA-DA) and Superpixelwise PCA (SuperPCA), we introduce SuperPCA based DA (SuperPCA-DA) in this paper. Specifically, an HSI is firstly divided into various superpixel blocks by typical image segmentation techniques, followed by fixed-size window and superpixel based local reconstruction for HSI denoising, and then the proposed SuperPCA-DA is employed in each superpixel block for HSI data augmentation where PCA is locally conducted in each superpixel to extract low-dimensional features, which will be projected back onto the original high-dimensional spectral space with random noises added into the projection matrix of PCA. The novel DA model can effectively generate new samples with diversity and richness by employing superpixels based spatial information of HSI, which could outperform classic and state-of-the-art DA models and then improve the accuracy of the subsequent classification models, especially when the number of training data is limited. The experimental results on three HSI datasets demonstrate the effectiveness of the proposed method. The code of the proposed model is available at https://github.com/XinweiJiang/SuperPCA-DA. [ABSTRACT FROM AUTHOR]

Hyperspectral Image Processing (HSIP) is an essential technique in remote sensing. Currently, extensive research is carried out in hyperspectral image processing, involving many applications, including land cover classification, anomaly detection, plant classification, etc., Hyperspectral image processing is a powerful tool that enables us to capture and analyze an object's spectral information with greater accuracy and precision. Hyperspectral images are made up of hundreds of spectral bands, capturing an immense amount of information about the earth's surface. Accurately classifying and predicting land cover in these images is critical to understanding our planet's ecosystem and the impact of human activities on it. With the advent of deep learning techniques, the process of analyzing hyperspectral images has become more efficient and accurate than ever before. These techniques enable us to categorize land cover and predict Land Use/Land Cover (LULC) with exceptional precision, providing valuable insights into the state of our planet's environment. Image classification is difficult in hyperspectral image processing because of the large number of data samples but with a limited label. By selecting the appropriate bands from the image, we can get the finest classification results and predicted values. To our knowledge, the previous review papers concentrated only on the classification method. Here, we have presented an extensive review of various components of hyperspectral image processing, hyperspectral image analysis, pre-processing of an image, feature extraction and feature selection methods to select the number of features (bands), classification methods, and prediction methods. In addition, we also elaborated on the datasets used for classification, evaluation metrics used, various issues, and challenges. Thus, this review article will benefit new researchers in the hyperspectral image classification domain. [ABSTRACT FROM AUTHOR]

As technology progresses, automation gains importance. The automation might be in a large-scale industry with more employees and heavy capital investments, or maybe a small-scale industry in which manufacturing, providing the services, and productions are performed on a smaller scale or micro scale, everywhere the automation gains importance. Similar to this is the food processing industry, where fruits and vegetables are some of the most popular products that enhance our health and help us to stay fit. In this approach, we have developed a framework that classifies fruits and vegetables using computer vision and deep learning-based methods. We test the proposed framework on Kaggle's fresh/stale images of fruits and vegetables dataset and IEEEDataPort's FruitsGB dataset. Experiments were conducted in multiple trials to extract model parameters and were analyzed to classify the fruits and vegetables as fresh/stale. The classification depends on the selection of the optimizer and varying the hyperparameter value like batch size, learning rate, kernel size, number of kernels, patch size, etc. The proposed custom CNN model achieves the highest classification accuracy of 97.65% and 95.86% using Kaggle's and FruitsGB datasets, respectively. Similarly, in the second approach, the vision transform (ViT) achieves the highest classification accuracy of 98.34% and 96.75% on the same datasets, respectively. The results of these methods outperform the results of the baseline algorithm used in the classification of the images. [ABSTRACT FROM AUTHOR]

The rapid development of modern society and continuous urbanization have resulted in a proliferation of functional buildings, which offer significant convenience to individuals, but pose significant fire hazards as well. How to detect the fire at the early stage is always the focus of research. This paper proposes a multi-information source fusion fire recognition method based on particle swarm optimization (PSO)-backpropagation (BP) neural networks and ResNet50. The PSO algorithm is applied to optimize the initial parameters of a BP neural network model, while data from three sensors — temperature, humidity and smoke — are integrated, through iterative training of the system, accurate recognition of sensor data can be achieved. Additionally, a method is proposed for the recognition of infrared fire images using ResNet50 and transfer learning. By improving the ResNet50 network model and migrating the ResNet50 pre-trained network weight, infrared fire image recognition accuracy is further enhanced. Then the sensor information recognition results and image information recognition results are input into the fuzzy system for fusion reasoning again, and the final decision is output according to the set fuzzy rules. Experimental findings demonstrate that the multi-information source fusion approach utilizing the PSO-BP neural network and ResNet50 significantly enhances the accuracy and response time of fire recognition, and achieves a remarkable recognition effect. [ABSTRACT FROM AUTHOR]

This paper presents a study on the computational complexity of coding for machines, with a focus on image coding for classification. We first conduct a comprehensive set of experiments to analyze the size of the encoder (which encodes images to bitstreams), the size of the decoder (which decodes bitstreams and predicts class labels), and their impact on the rate–accuracy trade-off in compression for classification. Through empirical investigation, we demonstrate a complementary relationship between the encoder size and the decoder size, i.e., it is better to employ a large encoder with a small decoder and vice versa. Motivated by this relationship, we introduce a feature compression-based method for efficient image compression for classification. By compressing features at various layers of a neural network-based image classification model, our method achieves adjustable rate, accuracy, and encoder (or decoder) size using a single model. Experimental results on ImageNet classification show that our method achieves competitive results with existing methods while being much more flexible. The code will be made publicly available. [ABSTRACT FROM AUTHOR]

This study attempts to classify ear morphology for human identification in forensic investigations by distinguishing between the upper auricle and lobule areas. A three-dimensional homologous model of the ear was created using 414 ear images of males aged 17–93 years reconstructed from computed tomography scans of forensic autopsy cases. Morphological changes were visualized using principal component analysis and areas of significant individual differences within the entire ear were identified. The classification criterion images for the upper auricle (ten images) and lobule (12 images) were developed by combining multiple principal component values: components 1–5 for the upper auricle and 1–6 for the lobule. Three-dimensional ear images of the upper auricle and lobule areas from 414 subjects were categorized using a measurement method based on the minimum distance between 5,507 corresponding points. The results indicate the applicability of the criterion images for the morphological classification of ears in this study. [ABSTRACT FROM AUTHOR]

Mobile Ru ions in oxide media have been reported as a novel species that offer extremely low switching currents for memristors. However, their bi‐stable resistive‐switching (RS) and low‐switching currents dynamics have not been quantitatively unveiled. Here, the bi‐stable RS mechanism via in‐depth field‐induced atomic migration and chemical bonding state studies is elucidated, showing that the RS of the Ru‐based hybrid memristor (RHM) is possible via the simultaneously controlled hybrid Ru cation and oxygen anion. Additionally, the Ru ion mobility is quantitatively obtained via atomic moving distance and switching time measurements, demonstrating that the lower Ru ion mobility, compared to other conventional mobile species in oxide media, can be the origin of the low‐switching currents. It is found that the current conduction mechanism of the low‐resistance‐state in RHMs has temperature‐range‐dependencies. The direct tunneling conduction mechanism is dominant in relatively low temperatures; however, the ionic transport and thermally activated hopping conduction mechanism govern the current flow in high temperatures. Owing to the low Ru ion mobility, the RHM exhibits highly linear synaptic plasticity with a low‐conductance regime, showing outstanding energy efficiency compared to other memristors in image recognition tasks. These findings can contribute to improving the feasibility of hyper‐scale synaptic cores consisting of RHMs. [ABSTRACT FROM AUTHOR]

A novel polarization‐sensitive artificial optoelectronic synapse based on β‐Ga2O3 single‐crystal is proposed in this work, featuring reconfigurable anisotropic vision. A series of polarization‐sensitive synaptic activities and polarization‐sensitive image recognition functions are successfully simulated using this device. The intriguing performance of this device, stems from the crystal anisotropy in β‐Ga2O3, which is confirmed through polarization Raman measurements and first‐principles theoretical calculations. Furthermore, a comprehensive analysis of the persistent photoelectric properties of the device unveils that the adaptability of the optoelectronic synapse device stems from the ionization and dissociation of oxygen vacancies. Ultimately, the device is utilized in the fields of image recognition and information encryption. A four‐layer artificial neural network with two hidden layers is constructed for recognizing handwritten digits. After training, the recognition accuracy reaches over 91.5% for both unpolarized and polarized light. Information encryption is achieved by controlling polarization states. The device enables data generation and encryption to be conducted on the same platform, mitigating exposure risks during transmission and significantly enhancing data security and confidentiality. This work presents new opportunities for future applications of polarization‐based perception systems. [ABSTRACT FROM AUTHOR]

Handwritten character recognition falls under the domain of image classification, which has been under research for years. But still, specific gaps need to be highlighted as offline handwritten character recognition (OHCR) with the limitation of the unstructured hierarchy of character classification. However, the idea is to make the machine recognize handwritten human characters. The language focused on in this research paper is English, using offline handwritten character recognition for identifying English characters. There are many publicly available datasets, of which EMNIST is the most challenging. The key idea of this research paper is to recommend a deep learning-based ELBP-CNN method to help recognize English characters. This research paper proposes a deep learning CovNet with feature extraction and novel local binary pattern-based approaches, LBP (AND, OR), that is tested and compared with renowned pre-trained models using transfer learning. These parametric settings address multiple issues and are finalized after experimentation. The same hyperparametric settings were used for all the models under test and E-Character, with the same data augmentation settings. The proposed model, named the E-Character recognizer, produced 87.31% accuracy. It was better than most of the tested pre-trained models and other proposed methods by other researchers. This research paper further highlighted some problems, like misclassification due to the similar structure of characters. [ABSTRACT FROM AUTHOR]

The development and application of artificial intelligence-based computer vision systems in medicine, environment, and industry are playing an increasingly prominent role. Hence, the need for optimal and efficient hyperparameter tuning strategies is more than crucial to deliver the highest performance of the deep learning networks in large and demanding datasets. In our study, we have developed and evaluated a new training methodology named deep multi-metric training (DMMT) for enhanced training performance. The DMMT delivers a state of robust learning for deep networks using a new important criterion of multi-metric performance evaluation. We have tested the DMMT methodology in multi-class (three, four, and ten), multi-vendors (different X-ray imaging devices), and multi-size (large, medium, and small) datasets. The validity of the DMMT methodology has been tested in three different classification problems: (i) medical disease classification, (ii) environmental classification, and (iii) ecological classification. For disease classification, we have used two large COVID-19 chest X-rays datasets, namely the BIMCV COVID-19+ and Sheffield hospital datasets. The environmental application is related to the classification of weather images in cloudy, rainy, shine or sunrise conditions. The ecological classification task involves a classification of three animal species (cat, dog, wild) and a classification of ten animals and transportation vehicles categories (CIFAR-10). We have used state-of-the-art networks of DenseNet-121, ResNet-50, VGG-16, VGG-19, and DenResCov-19 (DenRes-131) to verify that our novel methodology is applicable in a variety of different deep learning networks. To the best of our knowledge, this is the first work that proposes a training methodology to deliver robust learning, over a variety of deep learning networks and multi-field classification problems. [ABSTRACT FROM AUTHOR]

Accurate information retrieval from multi-source and multi-resolution image data constitutes a foundation for knowledge discovery. Scene image classification in the remote sensing (RS) community using aerial very high resolution (VHR) images is one of the well-researched areas, which mostly utilise deep learning (DL)—based methods thanks to their remarkable classification performance. Nevertheless, existing DL-based methods still have a limited ability to capture precise spatial semantic information scattered toward the horizontal and vertical directions across such images at multiple scales and rotations. As such, we herein propose a novel approach, employing an innovative rotation invariant horizontal vertical pooled module (RIHVPM), to well-represent aerial VHR RS images for stable and improved classification performance. Notably, the proposed RIHVPM benefits from the multiple tensor rotations coupled with attention-enabled multiscale horizontal and vertical pooling operations for image representation. An experimental study on three benchmark datasets demonstrates competent and/or higher classification performance (AID: 96.44%, NWPU: 94.32% and UCM: 99.04%) and robustness/stability (minimum standard deviation of 0.001) of the proposed approach. [ABSTRACT FROM AUTHOR]

Automatically and accurately identifying the deformation zone of coastal slope landslides is crucial for exploring the mechanism of landslides and predicting landslide disasters. To this end, this study proposes an integrated automatic recognition method combining Image Clipping (IC), Image Information Enhancement (IE), Adaptive K-means Clustering Segmentation (AKS), and Optimization (O): IC-IE-AKS-O, which achieves precise extraction of the deformation area in coastal slope landslide images. Firstly, due to the more complex natural environment of field slopes, to extend the monitoring duration, we introduce a hierarchical operation algorithm based on the HSV color model, which effectively mitigates the impact of sunlight, rain, and foggy weather on image recognition accuracy. Secondly, this study proposes a 2D landslide image segmentation technique that combines K-means clustering with global threshold segmentation for landslide images, enabling the segmentation of small image regions with precision. Finally, we combine image information enhancement technology with image segmentation technology. To verify its effectiveness, we identify a landslide image of a coastal slope in Pingtan. The method displays an average relative error of 5.20% and 5.14% in the X and Y directions, respectively. Its advantages are threefold: (1) The combination of image information enhancement and segmentation techniques can more accurately identify landslide areas that appear blurred in the image; (2) expanding the temporal dimension of coastal slope monitoring; (3) providing excellent boundary conditions and segmentation results. The practical application of this method ensures the stable and accurate operation of the coastal slope monitoring system, providing a safeguard for the sustainable development of marine safety. [ABSTRACT FROM AUTHOR]

With its increasing global prevalence, lung cancer remains a critical health concern. Despite the advancement of screening programs, patient selection and risk stratification pose significant challenges. This study addresses the pressing need for early detection through a novel diagnostic approach that leverages innovative image processing techniques. The urgency of early lung cancer detection is emphasized by its alarming growth worldwide. While computed tomography (CT) surpasses traditional X-ray methods, a comprehensive diagnosis requires a combination of imaging techniques. This research introduces an advanced diagnostic tool implemented through image processing methodologies. The methodology commences with histogram equalization, a crucial step in artifact removal from CT images sourced from a medical database. Accurate lung CT image segmentation, which is vital for cancer diagnosis, follows. The Otsu thresholding method and optimization, employing Colliding Bodies Optimization (CBO), enhance the precision of the segmentation process. A local binary pattern (LBP) is deployed for feature extraction, enabling the identification of nodule sizes and precise locations. The resulting image underwent classification using the densely connected CNN (DenseNet) deep learning algorithm, which effectively distinguished between benign and malignant tumors. The proposed CBO+DenseNet CNN exhibits remarkable performance improvements over traditional methods. Notable enhancements in accuracy (98.17%), specificity (97.32%), precision (97.46%), and recall (97.89%) are observed, as evidenced by the results from the fractional randomized voting model (FRVM). These findings highlight the potential of the proposed model as an advanced diagnostic tool. Its improved metrics promise heightened accuracy in tumor classification and localization. The proposed model uniquely combines Colliding Bodies Optimization (CBO) with DenseNet CNN, enhancing segmentation and classification accuracy for lung cancer detection, setting it apart from traditional methods with superior performance metrics. [ABSTRACT FROM AUTHOR]

Marine and ocean pollution is one of the most serious environmental problems in the world. Marine plastics pose a significant threat to the marine ecosystem due to their negative effects. After passing through various processes, plastic waste accumulates on the seafloor and fragments into very small pieces known as microplastics. These microplastics are to blame for the extinction and death of aquatic life. This study obtained a hybrid underwater dataset containing 13,089 images, sized 300 × 300, including garbage and sea animals. In the proposed method, this dataset is used to develop our example projector deep feature generator. In this study, using the Resnet101 network in a sample projector build, the feature generator creates 6,000 features. Using NCA (Neighborhood Component Analysis), the best 1000 features from a pool of 6,000 are selected. The kNN (k-nearest neighbor) algorithm is then used to classify the resulting feature vectors. As validation techniques, both tenfold cross-validations were used. The hybrid dataset's best accuracy was calculated to be 99.35%. Our recommendation is successful based on the comparisons and calculated performance measures. [ABSTRACT FROM AUTHOR]

The trademark element recogniton is a crucial task in applications such as trademark brand evaluation and trademark infringement identification. In recent years, although modeling technology has made significant progress, small objects, similar objects, and objects with high conditional probability continue to be unable to be solved, due to the limitations of convolutional kernels. Based on semantic-aware region search and label dependency modeling, we propose a multi-input recognition framework for trademark elements (Mi-Tr) based on Transformer, which learns the complex dependencies between visual features and labels them through feature extraction using different convolutional networks and Transformer encoding. The proposed approach includes two visual feature-embedding modules that use modified VGG16 and ResNet101 as feature extractors to obtain feature information of trademark images in different dimensions. Simultaneously, the category labels are input into the transformer by embedding, using the order invariance of the transformer, thus, it is better to learn all types of dependencies between all features and labels. Additionally, the number of layers of the transformer and number of heads of the multiheaded attention were modified to find parameters that better match image features and label information. The experimental results on two datasets, METU and Logotypes of Different Companies, demonstrate that the classifier developed by our model performs significantly better in the multi-input classification of trademark image elements. [ABSTRACT FROM AUTHOR]

Fine-grained visual classification (FGVC) necessitates expert knowledge,which is expensive and requires a large training sample size. Consequently, using sample data acquired through the web has emerged as a novel approach for augmenting training samples. However, the web data often includes noisy samples, leading to misclassification of deep learning models. This paper presents a a meta-learning-base method called Data Reweighting Net (DR-Net). It enables the use of small, clean meta set as a guiding mechanism to accurately learn web image datasets that contain noise. More specifically, the DR-Net fully learns from small, clean meta set to discard noisy samples and obtain clean web samples through low similarity properties. DR-Net enables classification networks to adaptively learn training sets through sample weighting, mitigating the impact of noisy labels on classification learning. Our experiments on Web-bird, Web-aircraft, Web-car, CIFAR-10, and CIFAR-100 datasets demonstrate the feasibility of our proposed method. [ABSTRACT FROM AUTHOR]

Nowadays, unmanned aerial vehicles (UAVs) becomes more prominent because of their benefits namely versatile, inexpensive, scalable, and autonomous. The aerial image classification procedure has involved further attention to the UAV-allowed surveillance model facilitating objective detection, and tracking procedure and significantly enhancing the visual surveillance outcomes. Earlier methods to resolve the scene classification method dependent upon feature coding systems with unsupervised feature learning or low-level hand-engineered features can make mid-level image features with limited representational capabilities that usually limit them from the achievement of effectual outcomes. The development of convolutional neural networks (CNNs) helps to attain an efficient classification process in the area of image classification. This paper introduces a Pelican Optimization Algorithm with Convolutional-Recurrent Hop-Field Neural Network (POA-CRHFNN) model for the UAV image classification model. The presented POA-CRHFNN method initially involves preprocessing in two stages namely Gaussian Filter (GF) based noise elimination and contrast enhancement. In addition, the POA-CRHFNN method exploits the ShuffleNetv2 CNN model for feature vector generation. For aerial image classification, the CRHFNN model is used, which allocates different types of classes to the aerial images. The combination of Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) in the CRHFNN can capture both spatial and temporal dependencies in aerial images. The integration of the Hopfield Neural Network (HNN) can recognize and classify aerial images based on learned patterns and associations. Finally, the POA is applied for fine-tuning the hyperparameters of the CRHFNN model. By leveraging the behavioural patterns of pelicans, the POA may explore the search space effectively and potentially find better hyperparameter configurations for improved classification performance. The effectiveness of the POA-SCNN model can be investigated on a benchmark aerial image database and the simulation outcomes reported the higher outcome of the POA- CRHFNN method over existing techniques with maximum accuracy of 99.77%, precision of 98.01%, recall of 97.52%, and F-score of 97.69%. [ABSTRACT FROM AUTHOR]

Due to the potentially high cost and difficulty of obtaining labeled data in medical image classification, semi-supervised learning can effectively utilize limited labeled data and large amounts of unlabeled data to enhance the performance of classification models. However, most existing methods utilize pseudo-labeling techniques that filter out a large amount of data, which makes the unlabeled data underutilized. Moreover, the consistency regularization methods rely heavily on the perturbation function suffering from a lack of generalization ability. In this paper, we propose a semi-supervised medical image classification approach based on class prototype matching for soft pseudo-labels with consistency regularization. A class prototype matching module is used to predict soft pseudo-labels for unlabeled data, and the quality of the predicted pseudo-labels is improved by a meticulously designed cache queue with dynamic and unbiased updates. The linear mixture strategy is employed to handle both labeled and unlabeled data supplying the model with more complex data inputs, which boosts the ability of the model to learn intra- and inter-class features. Extensive experiments on ISIC2018 skin lesion dataset and the Chexpert chest disease dataset demonstrate that our approach has satisfactory performance and generalization ability on the medical image classification than other state-of-the-art semi-supervised methods. [ABSTRACT FROM AUTHOR]

The accurate classification of bone tumours is crucial for guiding clinical decisions regarding treatment and follow-up. However, differentiating between various tumour types is challenging due to the rarity of certain entities, high intra-class variability, and limited training data in clinical practice. This study proposes a multimodal deep learning model that integrates clinical metadata and X-ray imaging to improve the classification of primary bone tumours. The dataset comprises 1,785 radiographs from 804 patients collected between 2000 and 2020, including metadata such as age, affected bone site, tumour position, and gender. Ten tumour types were selected, with histopathology or tumour board decisions serving as the reference standard. Methods: Our model is based on the NesT image classification model and a multilayer perceptron with a joint fusion architecture. Descriptive statistics included incidence and percentage ratios for discrete parameters, and mean, standard deviation, median, and interquartile range for continuous parameters. Results: The mean age of the patients was 33.62 ± 18.60 years, with 54.73% being male. Our multimodal deep learning model achieved 69.7% accuracy in classifying primary bone tumours, outperforming the Vision Transformer model by five percentage points. SHAP values indicated that age had the most substantial influence among the considered metadata. Conclusion: The joint fusion approach developed in this study, integrating clinical metadata and imaging data, outperformed state-of-the-art models in classifying primary bone tumours. The use of SHAP values provided insights into the impact of different metadata on the model's performance, highlighting the significant role of age. This approach has potential implications for improving diagnostic accuracy and understanding the influence of clinical factors in tumour classification. [ABSTRACT FROM AUTHOR]

Spatial data mining is an important approach for collecting useful data from big datasets, especially remotely sensed images. This study tackles issues in environmental monitoring and management using sophisticated image processing. The Horse Herd Optimization-based VGG19 (HHO-VGG19) is proposed to improve land cover classification, recognition of objects, detection of changes, and detection of anomalies. The study used the BCDD dataset, which was scaled to 512 × 512 pixels, then applied Z-score normalization and extracted features using Principal Component Analysis (PCA). The VGG19 architecture was enhanced by utilizing Horse Herd Optimization to enhance image classification efficiency. The HHO-VGG19 model surpasses conventional techniques, with F1-score of 92%, a recall of 94%, an accuracy of 98.5%, and a 30-second execution time reduction. The findings indicate the efficiency of integrating sophisticated image processing with spatial data mining, giving an effective tool for remote sensing image processing in environmental uses including tracking ecosystems and handling of natural resources. [ABSTRACT FROM AUTHOR]

Techniques that exploit spectral-spatial information have proven to be very effective in hyperspectral image classification. Joint sparse representation classification (JSRC) is one such technique which has been extensively used for this purpose. However, the use of a single fixed-sized window has limited its ability to incorporate spatial information. Several techniques such as multiscale superpixels based sparse representation classification (MSSRC), multiscale adaptive sparse representation classification (MASRC) and Discriminant Subdictionary Learning (DSDL) have tried to overcome this drawback by fusing information from different scales. However, their inability to simultaneously consider the correlated information at different scales and appropriate spatial neighbourhoods limits their performance. In order to better model contextual information, in this paper, we propose a modified max-tree and modified min-tree to represent the connected components of the image. Then, by exploiting these connected components, adaptive multiscale windows are defined. The potentiality of the proposed technique is validated by performing a comparative analysis with four state-of-the-art sparse representation methods using three real hyperspectral datasets. For a fixed training and test sets of University of Pavia and Indian Pines dataset, our proposed technique provides at least 3% and 2%, respectively higher classification results than the best state-of-the-art method. [ABSTRACT FROM AUTHOR]

Image splicing is a common and informal way to generate forgery images. Two or more than two images merged into one image is called image splicing. Many researchers in the image splicing area continue to face the challenge of splice forgery image recognition. This is why it's crucial to regularly update image forgery detection systems so that they are continuously updated and require more research in this area. The IFC-TC, DVMM, Columbia, and CASIA datasets have been used to evaluate the splicing recognition result of the proposed technique. In this technique initially, the features are extracted by using the DWT and LBP techniques for computing different parameters. Initially, images undergo three layers of DWT, and splitting the image into low- and sub-bands high-frequency occurs after the first level. In addition, by splitting a low-frequency sub-band twice, more information can be extracted, resulting in a three-level DWT. The different parameters computed using LBP and DWT are PPV, TPR, FPR, F-1 score, and accuracy for splicing forgery detection. The different parameters derived from DWT and LBP techniques are compared individually as well as combined. The SVM classifier is used to distinguish among original and spliced image identification. The PPV, TPR, FPR, F-1 Score, and accuracy parameters are 98.88%, 97.59%, 87.51%, 98.53%, and 98.31%, respectively, for the IFC-TC dataset. The PPV, TPR, FPR, F-1 Score, and accuracy parameters are 97.92%, 98.65%, 89.69%, 98.20%, and 98.59%, respectively, for the DVMM dataset. The same PPV, TPR, FPR, F-1 Score, and accuracy parameters were calculated for the Columbia and CASIA datasets, and the results were compared using the existing technique. [ABSTRACT FROM AUTHOR]

Rice diseases and pests have posed a serious threat to agricultural production in recent years. Particularly, false and missed detections have been caused by the diversity, scale, and complex growth environments of these pests. This study aims to enhance the detection accuracy of rice diseases and pests, in order to facilitate the deployment of detection models on edge devices. An improved YOLOv8-based model was developed to detect rice diseases and pests. Several innovations were also incorporated to optimize the performance of the model. Firstly, partial convolution (PConv) was introduced to design the PCBlock structure and then integrated into the feature extraction module of YOLOv8. As such, the module of lightweight feature extraction known as C2f-PCBlock significantly reduced the number of model parameters, thereby enhancing the detection speed without compromising accuracy. The complexity of the model was then reduced to deploy on the resource- constrained edge devices. Secondly, the semantic information was diluted to cause the feature fusion across non-adjacent layers. A balanced Feature fusion layer was then added to the neck network. feature pyramid networks (FPN) was restructured into balanced feature pyramid (BFP). The preservation of critical features was improved during fusion. Additionally, the embedded Gaussian non-local attention (EGNA) mechanism was incorporated to mitigate the aliasing behavior from the multi- layer feature fusion. The important features were retained to reduce the loss of information, leading to more accurate detections. The loss function of the model was also enhanced to replace the standard IoU-based loss with MPDIOU. The distortions were avoided in the detection bounding boxes, due to the significant variations in sample sizes. MPDIoU was used to reduce these distortions for less computational burden, indicating more efficient training. Extensive experiments were conducted on a rice pest and disease dataset. The results demonstrate that the improved model, named PBM-YOLOv8, outperformed the rest. Specifically, PBM-YOLOV8 was achieved in the precision rate of 6.1, 8.0, 4.2, 1.3, and 1.9 percentage points higher than Fast R-CNN, SSD, YOLOV5n, YOLOv8n, and YOLOv9t, respectively. The recall rates were also significantly improved, with an increase of 7.1, 8.4, 5.6, 2.8, and 3.2 percentage points, respectively. Furthermore, the mean average precision (mAP) values were 5.7, 7.3, 3.8, 1.1, and 1.5 percentage points higher than those of the compared models, respectively. PBM-YOLOv8 was selected as a leading model, in terms of detection accuracy. Moreover, the PBM-YOLOv8 model was highly efficient with only 4.8M parameters, which was far less than the 41.3M of Fast R-CNN and the 15.6 M of SSD. A high processing speed was maintained with an FPS of about 92, in order to balance the accuracy with computational efficiency. The effectiveness of the improved model was further validated. PBM-YOLOv8 was deployed on the RK3588 embedded board. The deployment showed an accelerated detection speed of 71.4 frames per second, with an average accuracy of 95.1%. This performance can fully meet the demands of practical applications, thus enabling real-time and accurate detection of rice pests and diseases. The multi-thread optimization was adopted in the deployment. The parallel acceleration of multi-core NPU was also realized for the model inference and CPU for post-processing. The model has further enhanced the efficiency and feasibility for real-world use. [ABSTRACT FROM AUTHOR]

In recent years, digital dentistry has increasingly utilized advanced image analysis techniques, such as image classification and disease diagnosis, to improve clinical outcomes. Despite these advances, the lack of comprehensive benchmark datasets is a significant barrier. To address this gap, our research team develop LMCD-OR, a substantial collection of oral radiograph images designed to support extensive artificial intelligence (AI)-driven diagnostics. LMCD-OR comprises 3,818 digital imaging and communications in medicine (DICOM) oral X-ray images from local medical institutions that are meticulously annotated to provide broad category information for both primary dental outpatient services and detailed secondary disease diagnoses. This dataset is engineered to train and validate multiclassification models to improve the precision and scope of oral disease diagnostics. To ensure robust dataset validation, we employ four cutting-edge visual neural network classification models as benchmarks. These models are tested against rigorous performance metrics, demonstrating the ability of the dataset to support advanced image classification and disease diagnosis tasks. LMCD-OR is publicly available at http://dentaldataset.zeroacademy.net. [ABSTRACT FROM AUTHOR]

Multi‐level anticounterfeiting tags have been developed using a combination of different materials. Polyvinyl alcohol (PVA) mixed with titanium dioxide (TiO2) is used to produce flexible substrates. Fluorescent Opuntia Ficus‐indica (OFI) extract dissolved with polymethyl methacrylate (PMMA) is then sprayed over the substrate to create a random, yet unique deposition of droplets. Photographs of the tags are taken under UV illumination at different angles and analyzed through the scale‐invariant feature transform (SIFT) algorithm to extract their unique features. The SIFT analysis reveals hundreds to thousands of matched features when a given tag is compared with itself, whereas this number drops to tens for different tags. To enhance the security of the tags, ITO is sputtered onto one of them in the form of a pattern formed by a patch array exhibiting a specific fingerprint at terahertz (THz) frequencies. The evaluation of ITO reflectance shows that each patch array has a unique and unpredictable response stemming from its distinct electro‐optical characteristics. The non‐deterministic response of sprayed dye droplets and ITO patches enables the realization of two‐level authentication, which is difficult to replicate at a reasonable cost. The simple manufacturing process and inexpensive materials involved make the proposed tags easily integrable into packaging. [ABSTRACT FROM AUTHOR]

This paper presents a quantitative index introduced, ‘clarity,’ which can address both the ultrasound probe's brightness and full contact states to indicate ultrasound image quality for the robotic ultrasound system. Furthermore, it evaluates its performance through an experiment, a comparative analysis, and a questionnaire. The index is calculated based on pixel statistics within the predefined region of interest. Experimental validation is conducted by implementing a robotic ultrasound system and performing experiments using two phantoms. Notably, the Results demonstrate the distinction of the visible difference in ultrasound images but also show that it is hard to classify the tiny difference. Additionally, a questionnaire was carried out to ask about the recognition of good image quality. The opinions are the same trends and patterns. The 53.4 clarity value was determined from the answers as the boundary between good and bad-quality images. The experiment results, and questionnaire findings suggest that the proposed clarity index could be a reliable tool for evaluating ultrasound image quality in robotic systems, offering insights for improving diagnostic accuracy and efficiency. This research provides a clear framework for assessing ultrasound image quality in robotic applications, paving the way for advancements in the efficacy and precision of ultrasound imaging techniques. [ABSTRACT FROM AUTHOR]

Hyperspectral image (HSI) data has a wide range of valuable spectral information for numerous tasks. HSI data encounters challenges such as small training samples, scarcity, and redundant information. Researchers have introduced various research works to address these challenges. Convolution Neural Network (CNN) has gained significant success in the field of HSI classification. CNN's primary focus is to extract low-level features from HSI data, and it has a limited ability to detect long-range dependencies due to the confined filter size. In contrast, vision transformers exhibit great success in the HSI classification field due to the use of attention mechanisms to learn the long-range dependencies. As mentioned earlier, the primary issue with these models is that they require sufficient labeled training data. To address this challenge, we proposed a spectral-spatial feature extractor group attention transformer that consists of a multiscale feature extractor to extract low-level or shallow features. For high-level semantic feature extraction, we proposed a group attention mechanism. Our proposed model is evaluated using four publicly available HSI datasets, which are Indian Pines, Pavia University, Salinas, and the KSC dataset. Our proposed approach achieved the best classification results in terms of overall accuracy (OA), average accuracy (AA), and Kappa coefficient. As mentioned earlier, the proposed approach utilized only 5%, 1%, 1%, and 10% of the training samples from the publicly available four datasets. [ABSTRACT FROM AUTHOR]

As the demand for renewable energy sources increases, the vulnerability of solar panels to lightning strikes becomes a critical concern. This research explores the correlation between lightning-induced voltage fluctuations and the resultant damage intensity on solar panels. The study adopts a systematic approach, first investigating the correlation between lightning-induced voltage assessment using 30kV, 60kV and 90 kV impulse voltage with multi-stage Marx impulse generator and the damage intensity on monocrystalline and polycrystalline solar panels. A portable active infrared thermography equipment with tCam-Mini was employed to conduct this research with wireless streaming. Furthermore, the study integrates Convolutional Neural Network (CNN)-based image classification techniques to enhance the efficiency of damage assessment. The results highlight the potential of neural networks in improving the accuracy and speed of image classification for damaged and undamaged samples. The findings contribute valuable insights into enhancing the resilience of solar panel systems against lightning strikes, ultimately advancing the reliability and sustainability of solar energy infrastructure. A new CNN model was developed to classify the images obtained from thermography with 90.21% accuracy for greyscale and 85.31% accuracy on thermal images. [ABSTRACT FROM AUTHOR]

Overheating of power transformers, low-voltage panels, and medium-voltage components in Electric Transformer Rooms (ETRs) can result from various factors, such as contact issues, irregular loads, and other similar problems. Thermal imaging shows significant potential for detecting faults in power equipment. However, its effectiveness is hindered by the complex thermal patterns of faults and variability in equipment and environmental conditions, making accurate fault detection challenging. This paper aims to study the effectiveness of transfer learning architectures for automating equipment classification in ETRs. This work applies four transfer learning architectures: AlexNet, SqueezeNet, VGG19, and GoogLeNet. The findings of the testing phase demonstrated that the use of transfer learning by fine-tuning pre-trained convolutional neural networks was highly effective in the classification of thermal images captured from ETRs, with the models achieving accuracy rates between 86.98% and 100%, and F 1 -Scores between 86.79% and 100%. [Display omitted] • Transfer learning is applied to automate the classification of thermal images of equipment in ETRs. • A thermal imaging dataset is set up and augmented to improve the generalization abilities of proposed models. • Establish standard comparisons among four transfer learning architectures. [ABSTRACT FROM AUTHOR]

Accurate detection of tea leaf diseases and insects is crucial for their scientific and effective prevention and control, essential for ensuring the quality and yield of tea. Traditional methods for identifying tea leaf diseases and insects primarily rely on professional technicians, which are difficult to apply in various scenarios. This study proposes a recognition method for tea leaf diseases and insects based on improved MobileNetV3. Initially, a dataset containing images of 17 different types of tea leaf diseases and insects was curated, with data augmentation techniques utilized to broaden recognition scenarios. Subsequently, the network structure of MobileNetV3 was enhanced by integrating the CA (coordinate attention) module to improve the perception of location information. Moreover, a fine-tuning transfer learning strategy was employed to optimize model training and accelerate convergence. Experimental results on the constructed dataset reveal that the initial recognition accuracy of MobileNetV3 is 94.45%, with an F1-score of 94.12%. Without transfer learning, the recognition accuracy of MobileNetV3-CA reaches 94.58%, while with transfer learning, it reaches 95.88%. Through comparative experiments, this study compares the improved algorithm with the original MobileNetV3 model and other classical image classification models (ResNet18, AlexNet, VGG16, SqueezeNet, and ShuffleNetV2). The findings show that MobileNetV3-CA based on transfer learning achieves higher accuracy in identifying tea leaf diseases and insects. Finally, a tea diseases and insects identification application was developed based on this model. The model showed strong robustness and could provide a reliable reference for intelligent diagnosis of tea diseases and insects. [ABSTRACT FROM AUTHOR]

Introduction: Tomatoes are sensitive to low temperatures during their growth process, and low temperatures are one of the main environmental limitations affecting plant growth and development in Northeast China. Chlorophyll fluorescence imaging technology is a powerful tool for evaluating the efficiency of plant photosynthesis, which can detect and reflect the effects that plants are subjected to during the low temperature stress stage, including early chilling injury. Methods: This article primarily utilizes the chlorophyll fluorescence image set of tomato seedlings, applying the dung beetle optimization (DBO) algorithm to enhance the deep learning bidirectional long short term memory (BiLSTM) model, thereby improving the accuracy of classification prediction for chilling injury in tomatoes. Firstly, the proportion of tomato chilling injury areas in chlorophyll fluorescence images was calculated using a threshold segmentation algorithm to classify tomato cold damage into four categories. Then, the features of each type of cold damage image were filtered using SRCC to extract the data with the highest correlation with cold damage. These data served as the training and testing sample set for the BiLSTM model. Finally, DBO algorithm was applied to enhance the deep learning BiLSTM model, and the DBO-BiLSTM model was proposed to improve the prediction performance of tomato seedling category labels. Results: The results showed that the DBO-BiLSTM model optimized by DBO achieved an accuracy, precision, recall, and F1 score with an average of over 95%. Discussion: Compared to the original BiLSTM model, these evaluation parameters improved by 9.09%, 7.02%, 9.16%, and 8.68%, respectively. When compared to the commonly used SVM classification model, the evaluation parameters showed an increase of 6.35%, 7.33%, 6.33%, and 6.5%, respectively. This study was expected to detect early chilling injury through chlorophyll fluorescence imaging, achieve automatic classification and labeling of cold damage data, and lay a research foundation for in-depth research on the cold damage resistance of plants themselves and exploring the application of deep learning classification methods in precision agriculture. [ABSTRACT FROM AUTHOR]

Introduction: Colorectal cancer (CRC) is one of the main causes of deaths worldwide. Early detection and diagnosis of its precursor lesion, the polyp, is key to reduce its mortality and to improve procedure efficiency. During the last two decades, several computational methods have been proposed to assist clinicians in detection, segmentation and classification tasks but the lack of a common public validation framework makes it difficult to determine which of them is ready to be deployed in the exploration room. Methods: This study presents a complete validation framework and we compare several methodologies for each of the polyp characterization tasks. Results: Results show that the majority of the approaches are able to provide good performance for the detection and segmentation task, but that there is room for improvement regarding polyp classification. Discussion: While studied show promising results in the assistance of polyp detection and segmentation tasks, further research should be done in classification task to obtain reliable results to assist the clinicians during the procedure. The presented framework provides a standarized method for evaluating and comparing different approaches, which could facilitate the identification of clinically prepared assisting methods. [ABSTRACT FROM AUTHOR]

Background: White matter hyperintensities (WMH) observed in T2 fluidattenuated inversion recovery (FLAIR) images have emerged as potential markers of neurodegenerative diseases like Multiple Sclerosis (MS). Lacking comprehensive automated WMH classification systems in current research, there is a need to develop accurate detection and classification methods for WMH that will benefit the diagnosis and monitoring of brain diseases. Objective: Juxtacortical WMH (JCWMH) is a less explored subtype of WMH, primarily due to the hard definition of the cortex in FLAIR images, which is escalated by the presence of lesions to obtain appropriate gray matter (GM) masks. Methods: In this study, we present a method to perform a specialized GM segmentation developed for the classification of WMH, especially JCWMH. Using T1 and FLAIR images, we propose a pipeline to integrate masks of white matter, cerebrospinal fluid, ventricles, and WMH to create a unique mask to refine the primary GM map. Subsequently, we utilize this pipeline to generate paired data for training a conditional generative adversarial network (cGAN) to substitute the pipeline and reduce the inputs to only FLAIR images. The classification of WMH is then based on the distances between WMH and ventricular and GM masks. Due to the lack of multi-class labeled WMH datasets and the need for extensive data for training deep learning models, we attempted to collect a large local dataset and manually segment and label some data for WMH and ventricles. Results: In JCWMH classification, the proposed method exhibited a Dice similarity coefficient, precision, and sensitivity of 0.76, 0.69, and 0.84, respectively. With values of 0.66, 0.55, and 0.81, the proposed method clearly outperformed the approach commonly used in the literature, which uses extracted GM masks from registered T1 images on FLAIR. Conclusion: After training, the method proves its efficiency by providing results in less than one second. In contrast, the usual approach would require at least two minutes for registration and segmentation alone. The proposed method is automated and fast and requires no initialization as it works exclusively with FLAIR images. Such innovative methods will undoubtedly facilitate accurate and meaningful analysis of WMH in clinical practice by reducing complexity and increasing efficiency. [ABSTRACT FROM AUTHOR]

Medical Object Detection (MOD) is a clinically relevant image processing method that locates structures of interest in radiological image data at object-level using bounding boxes. High-performing MOD models necessitate large datasets accurately reflecting the feature distribution of the corresponding problem domain. However, strict privacy regulations protecting patient data often hinder data consolidation, negatively affecting the performance and generalization of MOD models. Federated Learning (FL) offers a solution by enabling model training while the data remain at its original source institution. While existing FL solutions for medical image classification and segmentation demonstrate promising performance, FL for MOD remains largely unexplored. Motivated by this lack of technical solutions, we present an open-source, self-configuring and task-agnostic federated MOD framework. It integrates the FL framework Flower with nnDetection, a state-of-the-art MOD framework and provides several FL aggregation strategies. Furthermore, we evaluate model performance by creating simulated Independent Identically Distributed (IID) and non-IID scenarios, utilizing the publicly available datasets. Additionally, a detailed analysis of the distributions and characteristics of these datasets offers insights into how they can impact performance. Our framework's implementation demonstrates the feasibility of federated self-configuring MOD in non-IID scenarios and facilitates the development of MOD models trained on large distributed databases. [ABSTRACT FROM AUTHOR]

Automatic pavement disease detection aims to address the inefficiency in practical detection. However, traditional methods heavily rely on low-level image analysis, handcrafted features, and classical classifiers, leading to limited effectiveness and poor generalization in complex scenarios. Although significant progress has been made with deep learning methods, challenges persist in handling high-resolution images and diverse disease types. Therefore, this paper proposes a novel approach based on the lightweight Transformer Patch Labeling Network (LTPLN) to enhance the efficiency of automatic pavement disease detection and overcome the limitations of existing methods. Firstly, the input images undergo histogram equalization preprocessing to enhance image quality. Subsequently, the images are evenly partitioned into small patch blocks, serving as inputs to the enhanced Transformer model. This enhancement strategy involves integrating feature map labels at each layer of the model to reduce computational complexity and enhance model lightweightness. Furthermore, a depthwise separable convolution module is introduced into the Transformer architecture to introduce convolutional bias and reduce the model's dependence on large amounts of data. Finally, an iterative training process utilizing the label distillation strategy based on expectation maximization is employed to update the labels of patch blocks and roughly locate the positions of pavement diseases under weak supervision. Experimental results demonstrate that compared to the baseline model, the proposed enhanced model achieves a reduction of 2.5G Flops computational complexity and a 16% speed improvement on a private pavement disease dataset, with only a 1.2 percentage point decrease in AUC accuracy. Moreover, compared to other mainstream image classification models, this model exhibits more balanced performance on a public dataset, with improved accuracy and speed that better align with the practical requirements of pavement inspection. These findings highlight the significant performance advantages of the LTPLN model in automatic pavement disease detection tasks, making it more efficiently applicable in real-world scenarios. [ABSTRACT FROM AUTHOR]

Early screening of breast cancer through image recognition technology can significantly increase the survival rate of patients. Therefore, breast cancer pathological image is of great significance for medical diagnosis and clinical research. In recent years, numerous deep learning models have been applied to breast cancer image classification, with deep CNN being a typical representative. Due to the use of multi-depth small convolutional kernels in mainstream CNN architectures such as VGG and Inception, the obtained image features often have high dimensionality. Although high dimensionality can bring more fine-grained features, it also increases the computational complexity of subsequent classifiers and may even lead to the curse of dimensionality and overfitting. To address these issues, a novel embedded kernel CNN principal component feature fusion (CNN-PCFF) algorithm is proposed. The constructed kernel function is embedded in the principal component analysis to form the multi-kernel principal component. Multi-kernel principal component analysis is used to fuse the high dimensional features obtained from the convolution base into some representative comprehensive variables, which are called kernel principal components, so as to achieve the purpose of dimensionality reduction. Any type of classifier can be added based on multi-kernel principal components. Through experimental analysis on two public breast cancer image datasets, the results show that the proposed algorithm can improve the performance of the current mainstream CNN architecture and subsequent classifiers. Therefore, the proposed algorithm in this paper is an effective tool for the classification of breast cancer pathological images. [ABSTRACT FROM AUTHOR]

Currently, image recognition based on deep neural networks has become the mainstream direction of research; therefore, significant progress has been made in its application in the field of tea detection. Many deep models exhibit high recognition rates in tea leaves detection. However, deploying these models directly on tea-picking equipment in natural environments is impractical; the extremely high parameters and computational complexity of these models make it challenging to perform real-time tea leaves detection. Meanwhile, lightweight models struggle to achieve competitive detection accuracy; therefore, this paper addresses the issue of computational resource constraints in remote mountain areas and proposes Reconstructed Feature and Dual Distillation (RFDD) to enhance the detection capability of lightweight models for tea leaves. In our method, the Reconstructed Feature selectively masks the feature of the student model based on the spatial attention map of the teacher model; it utilizes a generation block to force the student model to generate the teacher's full feature. The Dual Distillation comprises Decoupled Distillation and Global Distillation. Decoupled Distillation divides the reconstructed feature into foreground and background features based on the Ground-Truth. This compels the student model to allocate different attention to foreground and background, focusing on their critical pixels and channels. However, Decoupled Distillation leads to the loss of relation knowledge between foreground and background pixels. Therefore, we further perform Global Distillation to extract this lost knowledge. Since RFDD only requires loss calculation on feature map, it can be easily applied to various detectors. We conducted experiments on detectors with different frameworks, using a tea dataset collected at the Huangshan Houkui Tea Plantation. The experimental results indicate that, under the guidance of RFDD, the student detectors have achieved performance improvements to varying degrees. For instance, a one-stage detector like RetinaNet (ResNet-50) experienced a 3.14% increase in Average Precision (AP) after RFDD guidance. Similarly, a two-stage model like Faster RCNN (ResNet-50) obtained a 3.53% improvement in AP. This offers promising prospects for lightweight models to efficiently perform real-time tea leaves detection tasks. [ABSTRACT FROM AUTHOR]