Search

Abstract SAR-to-optical (S2O) translation is able to covert SAR into optical images, which help the interpreter to extract information efficiently. In the absence of strictly matched datasets, it is difficult for existing methods to complete training on unpaired data with a minimum amount of data. By employing the recent Schrödinger bridge-based transformation framework, a multiscale axial residual module (MARM) based on the concept of multi-scale feature fusion has been proposed in this paper. To enable efficient translation of SAR to optical images, the generator and discriminator of the model have been designed. Extensive experiments on the SEN1-2 dataset conducted, and the results show the superiority of the proposed method in terms of the generation quality. Compared with the classical CycleGAN, the proposed method can improve the FID metrics by 42.05%.

In poor lighting and rainy and foggy bad weather environments, road traffic signs are blurred and have low recognition, etc. A super-resolution reconstruction algorithm for complex lighting and bad weather traffic sign images was proposed. First, a novel attention residual module was designed to incorporate an aggregated feature attention mechanism on the jump connection side of the base residual module so that the deep network can obtain richer detail information; second, a cross-layer jump connection feature fusion mechanism was adopted to enhance the flow of information across layers as well as to prevent the problem of gradient disappearance of the deep network to enhance the reconstruction of the edge detail information; and lastly, a positive-inverse dual-channel sub-pixel convolutional up-sampling method was designed to reconstruct super-resolution images to obtain better pixel and spatial information expression. The evaluation model was trained on the Chinese traffic sign dataset in a natural scene, and when the scaling factor is 4, the average values of PSNR and SSIM are improved by 0.031 when compared with the latest release of the deep learning-based super-resolution reconstruction algorithm for single-frame images, MICU (Multi-level Information Compensation and U-net), the average values of PSNR and SSIM are improved by 0.031 dB and 0.083, and the actual test average reaches 20.946 dB and 0.656. The experimental results show that the reconstructed image quality of this paper's algorithm is better than the mainstream algorithms of comparison in terms of objective indexes and subjective feelings. The super-resolution reconstructed image has a higher peak signal-to-noise ratio and perceptual similarity. It can provide certain technical support for the research of safe driving assistive devices in natural scenes under multi-temporal varying illumination conditions and bad weather.

Glaucoma is a common eye disease that can cause blindness. Accurate detection of the optic disc and cup disc is crucial for glaucoma diagnosis. Algorithm models based on artificial intelligence can assist doctors in improving detection performance. In this article, U-Net is used as the backbone network, and the attention and residual modules are integrated to construct an end-to-end convolutional neural network model for optic disc and cup disc segmentation. The U-Net backbone is used to infer the basic position information of optic disc and cup disc, the attention module enhances the model’s ability to represent and extract features of optic disc and cup disc, and the residual module alleviates gradient disappearance or explosion that may occur during feature representation of the neural network. The proposed model is trained and tested on the DRISHTI-GS1 dataset. Results show that compared with the original U-Net method, our model can more effectively separate optic disc and cup disc in terms of overlap error, sensitivity, and specificity.

UAVs （unmanned aerial vehicles） are now involved in intelligent transmission line inspection. Given the complex image backgrounds captured by UAVs and poor line inspection accuracy and low detection speed of UAVs， the paper proposes a power line inspection algorithm based on an improved SegNet model. Firstly， residual modules and asymmetric convolutions are introduced into the encoder to reduce the computational burden on the network. Secondly， the network layers of the decoding layer are reduced， and the features of the encoder and decoder are fused to improve inspection accuracy. Finally， the improved SegNet algorithm is used to train the power line dataset. The accuracy and mean intersection over union reach up to 89.4% and 86.62% respectively， and the single detection time is 46 ms. The experimental results show that the algorithm based on the improved SegNet model can achieve high-precision and real-time power line detection.

To solve the problem of missing data features using a deep convolutional neural network (DCNN), this paper proposes an improved gesture recognition method. The method first extracts the time-frequency spectrogram of surface electromyography (sEMG) using the continuous wavelet transform. Then, the Spatial Attention Module (SAM) is introduced to construct the DCNN-SAM model. The residual module is embedded to improve the feature representation of relevant regions, and reduces the problem of missing features. Finally, experiments with 10 different gestures are done for verification. The results validate that the recognition accuracy of the improved method is 96.1%. Compared with the DCNN, the accuracy is improved by about 6 percentage points.

Deep learning-based 3D target detection methods need to solve the problem of insufficient 3D target detection accuracy. In this paper, the KM3D network is selected as the benchmark network after the experimental comparison of current mainstream algorithms, and the IAE-KM3D network algorithm based on the KM3D network is proposed. First, the Resnet V2 network is introduced, and the residual module is redesigned to improve the training capability of the new residual module with higher generalization. IBN NET is then introduced to carefully integrate instance normalization and batch normalization as building blocks to improve the model’s detection accuracy in hue- and brightness-changing scenarios without increasing time loss. Then, a parameter-free attention mechanism, Simam, is introduced to improve the detection accuracy of the model. After that, the elliptical Gaussian kernel is introduced to improve the algorithm’s ability to detect 3D targets. Finally, a new key point loss function is proposed to improve the algorithm’s ability to train. Experiments using the KITTI dataset conclude that the IAE-KM3D network model significantly improves detection accuracy and outperforms the KM3D algorithm regarding detection performance compared to the original KM3D network. The improvements for AP2D, AP3D, and APBEV are 5%, 12.5%, and 8.3%, respectively, and only a tiny amount of time loss and network parameters are added. Compared with other mainstream target detection algorithms, Monn3D, 3DOP, GS3D, and FQNet, the improved IAE-KM3D network in this paper significantly improves AP3D and APBEV, with fewer network parameters and shorter time consumption.

The typical occlusion of cherry tomatoes in the natural environment is one of the most critical factors affecting the accurate picking of cherry tomato picking robots. To recognize occluded cherry tomatoes accurately and efficiently using deep convolutional neural networks, a new occluded cherry tomato recognition model DSP-YOLOv7-CA is proposed. Firstly, images of cherry tomatoes with different degrees of occlusion are acquired, four occlusion areas and four occlusion methods are defined, and a cherry tomato dataset (TOSL) is constructed. Then, based on YOLOv7, the convolution module of the original residual edges was replaced with null residual edges, depth-separable convolutional layers were added, and jump connections were added to reuse feature information. Then, a depth-separable convolutional layer is added to the SPPF module with fewer parameters to replace the original SPPCSPC module to solve the problem of loss of small target information by different pooled residual layers. Finally, a coordinate attention mechanism (CA) layer is introduced at the critical position of the enhanced feature extraction network to strengthen the attention to the occluded cherry tomato. The experimental results show that the DSP-YOLOv7-CA model outperforms other target detection models, with an average detection accuracy (mAP) of 98.86%, and the number of model parameters is reduced from 37.62MB to 33.71MB, which is better on the actual detection of cherry tomatoes with less than 95% occlusion. Relatively average results were obtained on detecting cherry tomatoes with a shade level higher than 95%, but such cherry tomatoes were not targeted for picking. The DSP-YOLOv7-CA model can accurately recognize the occluded cherry tomatoes in the natural environment, providing an effective solution for accurately picking cherry tomato picking robots.

In the cross-view condition, the gait recognition rate caused by the vastly different gait silhouette maps is substantially reduced. To improve the accuracy of gait recognition under cross-view conditions, this paper proposes a cross-view gait recognition network model combining multi-scale feature residual module (MFRM) and self-attention (SA) mechanism based on Generative Adversarial Network (GAN). First, the local and global feature information in the input gait energy image is fully extracted using the MFRM. Then, the SA mechanism module is used to adjust the information of channel dimensions and capture the association between feature information and is introduced into both the generator and discriminator. Next, the model is trained using a two-channel network training strategy to avoid the pattern collapse problem during training. Finally, the generator and discriminator are optimized to improve the quality of the generated gait images. This paper conducts experiments using the CASIA-B and OU-MVLP public datasets. The experiments demonstrate that the MFRM can better obtain the local and global feature information of the images. The SA mechanism module can effectively establish global dependencies between features, so that the generated gait images have clearer and richer detail information. The average Rank-1 recognition accuracies of the results in this paper reach 91.1% and 97.8% on the two datasets respectively, which are both better than the current commonly used algorithms, indicating that the network model in this paper can well improve the gait recognition accuracy across perspectives.

Pose estimation has been a hot topic in the field of machine vision in recent years. Animals exist widely in nature, and the analysis of their shape and movement is important in many fields and industries. In the pose estimation task, to improve the detection accuracy, the existing models often need to consume a lot of computing and memory resources. Therefore, it is a key problem for the pose estimation methods to carry out a lightweight model and reduce the computational overhead on the premise of ensuring model accuracy. In this paper, we focus on the structure of the convolutional neural network in animal pose estimation, construct a lightweight and efficient stacked hourglass network model oriented to optimize the balance of model computation and accuracy, and implement the application algorithm design based on it. Aiming at the problem of large parameters in depthwise convolutional neural networks, a lightweight residual module is proposed, that is, based on the lightweight efficient channel attention improved conditional channel-weighted method (ICCW-Bottle), thereby reducing the weight of the network and obtaining the feature information of different scales. Given the problem that a large amount of feature information is easily lost after the network pooling operation, a lightweight dual-branch fusion module is proposed that fully integrates high-level semantic information and low-level detailed features under the condition of a small number of parameters. Finally, the same as the CC-SSL method: the model is trained jointly using synthetic and real animal datasets, but the CC-SSL method does not take into account the computational power of the model, which consumes a lot of time and memory to run. Through experiments, it is known that compared with the CC-SSL method, the PCK@0.05 of this method is increased by 5.5% on the TigDog dataset. The model in this paper reduces the number of parameters and calculations of the network while ensuring less information loss and model accuracy. The ablation experiment verifies the advancement and effectiveness of the overall network.

Nasopharyngeal carcinoma (NPC) is a malignant tumor that occurs in the wall of the nasopharyngeal cavity and is prevalent in Southern China, Southeast Asia, North Africa, and the Middle East. According to studies, NPC is one of the most common malignant tumors in Hainan, China, and it has the highest incidence rate among otorhinolaryngological malignancies. We proposed a new deep learning network model to improve the segmentation accuracy of the target region of nasopharyngeal cancer. Our model is based on the U-Net-based network, to which we add Dilated Convolution Module, Transformer Module, and Residual Module. The new deep learning network model can effectively solve the problem of restricted convolutional fields of perception and achieve global and local multi-scale feature fusion. In our experiments, the proposed network was trained and validated using 10-fold cross-validation based on the records of 300 clinical patients. The results of our network were evaluated using the dice similarity coefficient (DSC) and the average symmetric surface distance (ASSD). The DSC and ASSD values are 0.852 and 0.544 mm, respectively. With the effective combination of the Dilated Convolution Module, Transformer Module, and Residual Module, we significantly improved the segmentation performance of the target region of the NPC.

Surface defect classification is one of key points in the field of steel manufacturing. It remains challenging primarily due to the rare occurrence of defect samples and the similarity between different defects. In this paper, a multi-level semantic method based on residual adversarial learning with Wasserstein divergence is proposed to realize sample augmentation and automatic classification of various defects simultaneously. Firstly, the residual module is introduced into model structure of adversarial learning to optimize the network structure and effectively improve the quality of samples generated by model. By substituting original classification layer with multiple convolution layers in the network framework, the feature extraction capability of model is further strengthened, enhancing the classification performance of model. Secondly, in order to better capture different semantic information, we design a multi-level semantic extractor to extract rich and diverse semantic features from real-world images to efficiently guide sample generation. In addition, the Wasserstein divergence is introduced into the loss function to effectively solve the problem of unstable network training. Finally, high-quality defect samples can be generated through adversarial learning, effectively expanding the limited training samples for defect classification. The experimental results substantiate that our proposed method can not only generate high-quality defect samples, but also accurately achieve the classification of defect detection samples.

The symmetrical segmentation of brain tumor images is crucial for both clinical diagnosis and computer-aided prognosis. Traditional manual methods are not only asymmetrical in terms of efficiency but also prone to errors and lengthy processing. A significant barrier to the process is the complex interplay between the deep learning network for MRI brain tumor imaging and the harmonious compound of both local and global feature information, which can throw off the balance in segmentation accuracy. Addressing this asymmetry becomes essential for precise diagnosis. In answer to this challenge, we introduce a balanced, end-to-end solution for brain tumor segmentation, incorporating modifications that mirror the U-Net architecture, ensuring a harmonious flow of information. Beginning with symmetric enhancement of the visual quality of MRI brain images, we then apply a symmetrical residual structure. By replacing the convolutional modules in both the encoder and decoder sections with deep residual modules, we establish a balance that counters the vanishing gradient problem commonly faced when the network depth increases. Following this, a symmetrical threefold attention block is integrated. This addition ensures a balanced fusion of local and global image features, fine-tuning the network to symmetrically discern and learn essential image characteristics. This harmonious integration remarkably amplifies the network’s precision in segmenting MRI brain tumors. We further validate the equilibrium achieved by our proposed model using three brain tumor segmentation datasets and four metrics and by juxtaposing our model against 21 traditional and learning-based counterparts. The results confirm that our balanced approach significantly elevates performance in the segmentation of MRI brain tumor images without an asymmetrical increase in computational time.

The precise segmentation of tumor regions plays a pivotal role in the diagnosis and treatment of brain tumors. However, due to the variable location, size, and shape of brain tumors, the automatic segmentation of brain tumors is a relatively challenging application. Recently, U-Net related methods, which largely improve the segmentation accuracy of brain tumors, have become the mainstream of this task. Following merits of the 3D U-Net architecture, this work constructs a novel 3D U-Net model called SGEResU-Net to segment brain tumors. SGEResU-Net simultaneously embeds residual blocks and spatial group-wise enhance (SGE) attention blocks into a single 3D U-Net architecture, in which SGE attention blocks are employed to enhance the feature learning of semantic regions and reduce possible noise and interference with almost no extra parameters. Besides, the self-ensemble module is also utilized to improve the segmentation accuracy of brain tumors. Evaluation experiments on the Brain Tumor Segmentation (BraTS) Challenge 2020 and 2021 benchmarks demonstrate the effectiveness of the proposed SGEResU-Net for this medical application. Moreover, it achieves DSC values of 83.31, 91.64 and 86.85%, as well as Hausdorff distances (95%) of 19.278, 5.945 and 7.567 for the enhancing tumor, whole tumor, and tumor core on BraTS 2021 dataset, respectively.

Aiming at the real-time detection and classification of the growth period of crops in the current digital cultivation and regulation technology of facility agriculture, an improved YOLOv4 method for identifying the growth period of strawberries in a greenhouse environment was proposed. The attention mechanism into the Cross Stage Partial Residual (CSPRes) module of the YOLOv4 backbone network was introduced, and the target feature information of different growth periods of strawberries while reducing the interference of complex backgrounds was integrated, the detection accuracy while ensured real-time detection efficiency was improved. Took the smart facility strawberry in Yunnan province as the test object, the results showed that the detection accuracy (AP) of the YOLOv4-CBAM model during flowering, fruit expansion, green and mature period were 92.38%, 82.45%, 68.01% and 92.31%, respectively, the mean average precision (mAP) was 83.78%, the mean inetersection over union (mIoU) was 77.88%, and the detection time for a single image was 26.13 ms. Compared with the YOLOv4-SC model, mAP and mIoU were increased by 1.62% and 2.73%, respectively. Compared with the YOLOv4-SE model, mAP and mIOU increased by 4.81% and 3.46%, respectively. Compared with the YOLOv4 model, mAP and mIOU increased by 8.69% and 5.53%, respectively. As the attention mechanism was added to the improved YOLOv4 model, the amount of parameters increased, but the detection time of improved YOLOv4 models only slightly increased. At the same time, the number of fruit expansion period recognized by YOLOv4 was less than that of YOLOv4-CBAM, YOLOv4-SC and YOLOv4-SE, because the color characteristics of fruit expansion period were similar to those of leaf background, which made YOLOv4 recognition susceptible to leaf background interference, and added attention mechanism could reduce background information interference. YOLOv4-CBAM had higher confidence and number of identifications in identifying strawberry growth stages than YOLOv4-SC, YOLOv4-SE and YOLOv4 models, indicated that YOLOv4-CBAM model can extract more comprehensive and rich features and focus more on identifying targets, thereby improved detection accuracy. YOLOv4-CBAM model can meet the demand for real-time detection of strawberry growth period status.

The continuous development of deep learning improves target detection technology day by day. The current research focuses on improving the accuracy of target detection technology, resulting in the target detection model being too large. The number of parameters and detection speed of the target detection model are very important for the practical application of target detection technology in embedded systems. This article proposed a real-time target detection method based on a lightweight convolutional neural network to reduce the number of model parameters and improve the detection speed. In this article, the depthwise separable residual module is constructed by combining depthwise separable convolution and non–bottleneck-free residual module, and the depthwise separable residual module and depthwise separable convolution structure are used to replace the VGG backbone network in the SSD network for feature extraction of the target detection model to reduce parameter quantity and improve detection speed. At the same time, the convolution kernels of 1 × 3 and 3 × 1 are used to replace the standard convolution of 3 × 3 by adding the convolution kernels of 1 × 3 and 3 × 1, respectively, to obtain multiple detection feature graphs corresponding to SSD, and the real-time target detection model based on a lightweight convolutional neural network is established by integrating the information of multiple detection feature graphs. This article used the self-built target detection dataset in complex scenes for comparative experiments; the experimental results verify the effectiveness and superiority of the proposed method. The model is tested on video to verify the real-time performance of the model, and the model is deployed on the Android platform to verify the scalability of the model.

Tumor segmentation using magnetic resonance imaging (MRI) plays a significant role in assisting brain tumor diagnosis and treatment. Recently, U-Net architecture with its variants have become prevalent in the field of brain tumor segmentation. However, the existing U-Net models mainly exploit coarse first-order features for tumor segmentation, and they seldom consider the more powerful second-order statistics of deep features. Therefore, in this work, we aim to explore the effectiveness of second-order statistical features for brain tumor segmentation application, and further propose a novel second-order residual brain tumor segmentation network, i.e., SoResU-Net. SoResU-Net utilizes a number of second-order modules to replace the original skip connection operations, thus augmenting the series of transformation operations and increasing the non-linearity of the segmentation network. Extensive experimental results on the BraTS 2018 and BraTS 2019 datasets demonstrate that SoResU-Net outperforms its baseline, especially on core tumor and enhancing tumor segmentation, illuminating the effectiveness of second-order statistical features for the brain tumor segmentation application.

Deaf and hearing-impaired people always face communication barriers. Non-invasive surface electromyography (sEMG) sensor-based sign language recognition (SLR) technology can help them to better integrate into social life. Since the traditional tandem convolutional neural network (CNN) structure used in most CNN-based studies inadequately captures the features of the input data, we propose a novel inception architecture with a residual module and dilated convolution (IRDC-net) to enlarge the receptive fields and enrich the feature maps, applying it to SLR tasks for the first time. This work first transformed the time domain signal into a time–frequency domain using discrete Fourier transformation. Second, an IRDC-net was constructed to recognize ten Chinese sign language signs. Third, the tandem CNN networks VGG-net and ResNet-18 were compared with our proposed parallel structure network, IRDC-net. Finally, the public dataset Ninapro DB1 was utilized to verify the generalization performance of the IRDC-net. The results showed that after transforming the time domain sEMG signal into the time–frequency domain, the classification accuracy (acc) increased from 84.29% to 91.70% when using the IRDC-net on our sign language dataset. Furthermore, for the time–frequency information of the public dataset Ninapro DB1, the classification accuracy reached 89.82%; this value is higher than that achieved in other recent studies. As such, our findings contribute to research into SLR tasks and to improving deaf and hearing-impaired people’s daily lives.

The deep convolutional neural network has achieved great success in the Single Image Super-resolution task. It is obviously that among the well-known super-resolution methods, the deep learning-based algorithms show the most advanced performance. However, the most advanced algorithms currently use complex networks with a large number of parameters, which makes it difficult to apply deep learning algorithms on mobile devices. To solve this problem, we propose a lightweight dual-residual network(LDRN) for single image super-resolution, which has better reconstruction quality than most current advanced lightweight algorithms. Due to its fewer parameters and computational expense, real-time and mobile applications of our networks can be easily realized. On the basis of the residual module, we propose a new residual unit, which uses two depthwise separable (DW) convolution to obtain better balance between feature extraction capacity and lightweight performance. We further design a dual-stream residual block, which contains a multiplication branch and an addition branch. The dual-stream residual block can improve the reconstruction performance more effectively than expanding the network width. In addition, we also designed a new up-sampling module to simplify the previous up-sampling methods. Extensive experimental results show that our network has better reconstruction performance and lightweight performance than most existing state-of-the-art algorithms. Our code is available at https://github.com/Jiangyichun-cust/pytorch-LDRN.

Abstract How to effectively protect ancient murals has become an urgent and important problem. Digital image processing developments have made it possible to repair damaged murals to a certain extent. This study proposes a consistency-enhanced generative adversarial network (GAN) model to repair missing mural areas. First, the convolutional layer from a fully convolutional network (FCN) is used to extract deep image features; then, through deconvolution, the features are mapped to the size of the original image and the repaired image is output, thereby completing the regenerative network. Next, global and local discriminant networks are applied to determine whether the repaired mural image is “authentic” in terms of both the modified and unmodified areas. In adversarial learning, the generative and discriminant network models are optimized to better complete the mural repair. The network introduces a dilated convolution that increases the convolution kernel’s receptive field. Each network convolutional layer joins in the batch standardization (BN) process to accelerate network convergence and increase the number of network layers and adopts a residual module to avoid the vanishing gradient problem and further optimizing the network. Compared with existing mural restoration algorithms, the proposed algorithm increases the peak signal-to-noise ratio (PSNR) by an average of 6–8 dB and increases the structural similarity (SSIM) index by 0.08–0.12. From a visual perspective, this algorithm successfully complements mural images with complex textures and large missing areas; thus, it may contribute to digital restorations of ancient murals.

Compared with zero-shot learning (ZSL), the generalized zero-shot learning (GZSL) is more challenging since its test samples are taken from both seen and unseen classes. Most previous mapping-based methods perform well on ZSL, while their performance degrades on GZSL. To solve this problem, inspired by the ensemble learning, this paper proposes a model with cooperative coupled generative networks (CCGN). Firstly, to alleviate the hubness problem, the reverse visual feature space is taken as the embedding space, with the mapping achieved by a visual feature center generation network. To learn a proper visual representation of each class, we propose a coupled of generative networks, which cooperate with each other to synthesize a visual feature center template of the class. Secondly, to improve the generative ability of the coupled networks, we further employ a deeper network to generate. Meanwhile, to alleviate loss semantic information problem caused by multiple network layers, a residual module is employed. Thirdly, to mitigate overfitting and to increase scalability, an adversarial network is introduced to discriminate the generation of visual feature centers. Finally, a reconstruction network, which reverses the generation process, is employed to restrict the structural correlation between the generated visual feature center and the original semantic representation of each class. Extensive experiments on five benchmark datasets (AWA1, AWA2, CUB, SUN, APY) demonstrate that the proposed algorithm yields satisfactory results, as compared with the state-of-the-art methods.

Brain tumor segmentation technology plays a pivotal role in the process of diagnosis and treatment of MRI brain tumors. It helps doctors to locate and measure tumors, as well as develop treatment and rehabilitation strategies. Recently, MRI brain tumor segmentation methods based on U-Net architecture have become popular as they largely improve the segmentation accuracy by applying skip connection to combine high-level feature information and low-level feature information. Meanwhile, researchers have demonstrated that introducing attention mechanism into U-Net can enhance local feature expression and improve the performance of medical image segmentation. In this work, we aim to explore the effectiveness of a recent attention module called attention gate for brain tumor segmentation task, and a novel Attention Gate Residual U-Net model, i.e., AGResU-Net, is further presented. AGResU-Net integrates residual modules and attention gates with a primeval and single U-Net architecture, in which a series of attention gate units are added into the skip connection for highlighting salient feature information while disambiguating irrelevant and noisy feature responses. AGResU-Net not only extracts abundant semantic information to enhance the ability of feature learning, but also pays attention to the information of small-scale brain tumors. We extensively evaluate attention gate units on three authoritative MRI brain tumor benchmarks, i.e., BraTS 2017, BraTS 2018 and BraTS 2019. Experimental results illuminate that models with attention gate units, i.e., Attention Gate U-Net (AGU-Net) and AGResU-Net, outperform their baselines of U-Net and ResU-Net, respectively. In addition, AGResU-Net achieves competitive performance than the representative brain tumor segmentation methods.

Accurate segmentation of retinal vessels is a basic step in diabetic retinopathy (DR) detection. Most methods based on deep convolutional neural network (DCNN) have small receptive fields, and hence they are unable to capture global context information of larger regions, with difficult to identify pathological. The final segmented retina vessels contain more noise with low classification accuracy. Therefore, in this paper, we propose a DCNN structure named as D-Net. In the encoding phase, we reduced the loss of feature information by reducing the downsampling factor, which reduced the difficulty of tiny thin vessels segmentation. We use the combined dilated convolution to effectively enlarge the receptive field of the network and alleviate the “grid problem” that exists in the standard dilated convolution. In the proposed multi-scale information fusion module (MSIF), parallel convolution layers with different dilation rates are used, so that the model can obtain more dense feature information and better capture retinal vessel information of different sizes. In the decoding module, the skip layer connection is used to propagate context information to higher resolution layers, so as to prevent low-level information from passing the entire network structure. Finally, our method was verified on DRIVE, STARE, and CHASE dataset. The experimental results show that our network structure outperforms some state-of-art method, such as N4-fields, U-Net, and DRIU in terms of accuracy, sensitivity, specificity, and ${AUC_{ROC}}$ . Particularly, D-Net outperforms U-Net by 1.04 %, 1.23 %, and 2.79% in DRIVE, STARE, and CHASE dataset, respectively.

The semantic segmentation of high-resolution remote sensing images (HRRSIs) is a basic task for remote sensing image processing and has a wide range of applications. However, the abundant texture information and wide imaging range of HRRSIs lead to the complex distribution of ground objects and unclear boundaries, which bring huge challenges to the segmentation of HRRSIs. To solve this problem, in this paper we propose an improved squeeze and excitation residual network (SERNet), which integrates several squeeze and excitation residual modules (SERMs) and a refine attention module (RAM). The SERM can recalibrate feature responses adaptively by modeling the long-range dependencies in the channel and spatial dimensions, which enables effective information to be transmitted between the shallow and deep layers. The RAM pays attention to global features that are beneficial to segmentation results. Furthermore, the ISPRS datasets were processed to focus on the segmentation of vegetation categories and introduce Digital Surface Model (DSM) images to learn and integrate features to improve the segmentation accuracy of surface vegetation, which has certain prospects in the field of forestry applications. We conduct a set of comparative experiments on ISPRS Vaihingen and Potsdam datasets. The results verify the superior performance of the proposed SERNet.

Detection of infected kernels is important for Fusarium head blight (FHB) prevention and product quality assurance in wheat. In this study, Raman spectroscopy (RS) and deep learning networks were used for the determination of FHB-infected wheat kernels. First, the RS spectra of healthy, mild, and severe infection kernels were measured and spectral changes and band attribution were analyzed. Then, the Inception network was improved by residual and channel attention modules to develop the recognition models of FHB infection. The Inception–attention network produced the best determination with accuracies in training set, validation set, and prediction set of 97.13%, 91.49%, and 93.62%, among all models. The average feature map of the channel clarified the important information in feature extraction, itself required to clarify the decision-making strategy. Overall, RS and the Inception–attention network provide a noninvasive, rapid, and accurate determination of FHB-infected wheat kernels and are expected to be applied to other pathogens or diseases in various crops.

Remote sensing images have been widely applied in various industries; nevertheless, the resolution of such images is relatively low. Panchromatic sharpening (pan-sharpening) is a research focus in the image fusion domain of remote sensing. Pan-sharpening is used to generate high-resolution multispectral (HRMS) images making full use of low-resolution multispectral (LRMS) images and panchromatic (PAN) images. Traditional pan-sharpening has the problems of spectral distortion, ringing effect, and low resolution. The convolutional neural network (CNN) is gradually applied to pan-sharpening. Aiming at the aforementioned problems, we propose a distributed fusion framework based on residual CNN (RCNN), namely, RDFNet, which realizes the data fusion of three channels. It can make the most of the spectral information and spatial information of LRMS and PAN images. The proposed fusion network employs a distributed fusion architecture to make the best of the fusion outcome of the previous step in the fusion channel, so that the subsequent fusion acquires much more spectral and spatial information. Moreover, two feature extraction channels are used to extract the features of MS and PAN images respectively, using the residual module, and features of different scales are used for the fusion channel. In this way, spectral distortion and spatial information loss are reduced. Employing data from four different satellites to compare the proposed RDFNet, the results of the experiment show that the proposed RDFNet has superior performance in improving spatial resolution and preserving spectral information, and has good robustness and generalization in improving the fusion quality.

The accurate segmentation of the paraspinal muscle in Magnetic Resonance (MR) images is a critical step in the automated analysis of lumbar diseases such as chronic low back pain, disc herniation and lumbar spinal stenosis. However, the automatic segmentation of multifidus and erector spinae has not yet been achieved due to three unusual challenges: (1) the muscle boundary is unclear; (2) the gray histogram distribution of the target overlaps with the background; (3) the intra- and inter-patient shape is variable. We propose to tackle the problem of the automatic segmentation of paravertebral muscles using a deformed U-net consisting of two main modules: the residual module and the feature pyramid attention (FPA) module. The residual module can directly return the gradient while preserving the details of the image to make the model easier to train. The FPA module fuses different scales of context information and provides useful salient features for high-level feature maps. In this paper, 120 cases were used for experiments, which were provided and labeled by the spine surgery department of Shengjing Hospital of China Medical University. The experimental results show that the model can achieve higher predictive capability. The dice coefficient of the multifidus is as high as 0.949, and the Hausdorff distance is 4.62 mm. The dice coefficient of the erector spinae is 0.913 and the Hausdorff distance is 7.89 mm. The work of this paper will contribute to the development of an automatic measurement system for paraspinal muscles, which is of great significance for the treatment of spinal diseases.