Search

• A hyperspectral image classification model with an improved generative adversarial network is proposed as a solution to the problem that conventional hyperspectral image classification methods necessitate a large number of labelled samples for model training and have poor classification accuracy. • Compared with the original generator composed of 4 layers of deconvolution layers, this paper uses a 6-layer residual network composed of upsampling layers and convolution layers to replace the generator network structure to improve the generation ability of target data. • Compared with the original 5-layer convolutional layer discriminator, this paper uses an 18-layer residual convolutional network to replace the convolutional layer network structure of the discriminator and improve the feature extraction capability of the algorithm. Classification of hyperspectral images is an essential application of deep learning techniques. However, standard deep learning approaches require a large number of labelled samples for model training, and classification performance can be enhanced. In this study, we propose the use of residual generative adversarial networks for the classification of hyperspectral images based on deep learning Convolutional Neural Networks, a large number of labelled samples, and the high classification accuracy required for the classification task. The method is based on generative adversarial networks, with a 6-layer residual network containing up-sampling and convolutional layers replacing the inverse convolutional layer network structure of the generator to enhance the data generation capability, and an 18-layer residual convolutional network replacing the convolutional layer network structure of the discriminator to enhance the feature extraction capability. The hyperspectral image classification method with residual generation adversarial network enhances the information exchange between the shallow network and the deep network by adding a residual structure to the network, extracts the deep features of hyperspectral images and improves the accuracy of hyperspectral image classification. Extensive experiments on several benchmark hyperspectral datasets have shown that the method outperforms comparative methods by 0.7% to 22.3% on OA. [ABSTRACT FROM AUTHOR]

• Taking advantage of the fact that CatBoost does not require significant preprocessing of category features, the internal construction uses unique thermal coding for categorical features with small bases, enabling the model to automatically transform category-based features into numerical features. • The proposed model supports automatic finding of class feature combinations, which significantly extends the dimensionality of license plate hyperspectral image features by maximizing the relationship between different features, further improving the effectiveness and generalization ability of the model. • By comparing with traditional image processing methods, the proposed CatBoost-based license plate hyperspectral image classification model has better performance in processing applications. The rapid development of artificial intelligence technology has brought profound technical changes to many traditional industries, among which intelligent transportation has become a hot spot for development in the traditional transportation field. Remote sensing license plate image recognition technology is extensively emphasized in such domains as intelligent transportation and intelligent vehicle management. However, in the practical traffic environment, low visibility scenes caused by complex environmental factors such as rain, snow, haze and cloudy days influence the recognition and classification of license plates, while the distortion of license plate images that may be caused by irregular movements of vehicles bring challenges to license plate recognition classification. The CNN-CatBoost model proposed in this paper divides the license plate recognition classification into two stages. The first stage uses the excellent performance of convolutional neural network in processing image data to extract various license plate image features; the second stage uses the CatBoost module to further process the image feature data and finally obtain the remote sensing license plate information. The model achieves outstanding results in the experiments and has practical application value. Through comparison with other network models, the CNN-CatBoost model proposed in this paper has superior performance. [ABSTRACT FROM AUTHOR]

• In this paper, a novel multi-stage convolutional neural network is proposed to maximize the extraction of HSI spatial-spectral features to improve the HSI accuracy of classification, addressing the shortcoming that the depth-separable convolutional classification accuracy is lower than that of ordinary convolutional networks. • In this paper, the HSI feature maps generated by CNN were fed into the cross-attention mechanism to obtain two sets of attention weights in different directions, and the feature extraction capability of the model was effectively improved by fusing the two sets of weights. • The two sets of weights generated by the cross-attention mechanism are operated by both summation and maxima to improve the model HSI classification accuracy. Hyperspectral images (HSI) are more informative than other remote sensing techniques and hence frequently utilized in numerous domains. Convolutional neural network algorithm provides outstanding performance in image processing and currently has established as the main approach in the field of HSI classification. HSI contain numerous channels with various spatial and spectral feature information, and simultaneously contain massive redundant information, leading to dimensional explosion or gradient disappearance in the process of classification. A multilayer network model based on Exponential Linear Units-guided Depthwise Separable Convolution is proposed in this paper to extract both spectral and spatial features from HSI at scales ranging, and the feature maps are then fed into a cross-attention mechanism for weight allocation to improve local feature information and optimize computational resource allocation. Extensive experiments on three well-known hyperspectral datasets are conducted, and the results show that the proposed network model can accurately complete the required HSI classification operation and outperforms other well-established techniques in terms of computing efficiency. [ABSTRACT FROM AUTHOR]

Image degradation commonly occurs when camera translation or irregular movement causes motion blur. In this paper, we remove the effects of unknown camera shake from an image using variational Bayesian estimation and the statistical characteristics of natural image gradients. We then propose a method to reduce ringing effects for image restoration that uses cyclic boundaries. Finally, we use the Richardson-Lucy algorithm to restore an image. Experimental results demonstrate that the blind restoration algorithm that is proposed in this paper can effectively remove image blurring caused by camera shake. Our algorithm can effectively suppress ringing artifacts and eliminate the effects of noise while protecting both edge and detail information. In general, the proposed motion deblurring method is effective in reducing both noise and ringing effects to enhance the quality of motion-blurred image restoration, and we obtained restored images that were clearer and more reliable. [ABSTRACT FROM AUTHOR]

• We propose a more robust feature representation method named HLDR to enhance the ability of uniting intra-class features and separating inter-class features under limited labeled samples. • We propose an SLG module based on Spatio-Temporal constraint, which can generate higher quality labels for unlabeled data than those only constrained by continuity in time. • We design a TSA strategy following a unified curriculum learning way to select some more accurate pseudo-labels to enlarge the limited training set. Cooperating with the above HLDR and SLG modules, we can achieve sufficient preponderance over other HIC methods on three benchmark data sets, especially with limited training data. With the development of deep learning technique, many hyperspectral image classification (HIC) methods achieve great progress based on the application of convolution neural network (CNN). However, most of them still face small data cases. In this paper, we study the challenges of small data in HIC and propose a stronger HIC small data learning framework (SDLF). More specifically, a feature enhancement module named high-low dimension hybrid representation (HLDR) is designed to improve the feature differentiation power. To generate higher quality classification labels for unlabeled data, a smooth label generation module (SLG) is introduced based on Spatio-Temporal constraint. Then, we introduce a dynamically training sample amplification (TSA) under a unified curriculum learning strategy to add generated reliable labels into training samples. Finally, integrating HLDR in feature representation, SLG in label generation, and TSA in training amplification, the proposed SDLF framework is constructed, which can ensure correctness of HIC network. Experimental results demonstrate that our SDLF can not only learn robust feature representation to enhance the differentiation power but also select reliably generated labels to enrich the training data, which is useful for learning from small data. Also, our SDLF method achieves a superior performance with small labeled training data on three typical datasets. [ABSTRACT FROM AUTHOR]

• AdvMixup avoids assigning ill-labels to the hyperspectral virtual samples. • AdvMixup retains the semantic information from the mixed samples. • The adversarial regularizer ensures the authenticity of generated samples. • AdvMixup helps to find an optimal hyperplane away from the high-density region. MixUp generates virtual samples by performing the identical convex combination in feature and semantic spaces simultaneously. However, convexly combining random pairs of images and their associated labels is likely to result in misunderstandings between the generated sample and its assigned pseudo label, particularly for hyperspectral images (HSI), leading to a confusing effect on the classifier. In this paper, adversarial MixUp with implicit semantic preservation (AdvMixup) is proposed for semi-supervised HSI classification (HSIc). The proposed AdvMixup exploits interpolation in the latent space for unlabeled samples. The virtual hyperspectral sample is generated by a deconvolutional network on the interpolated latent data. A semantic preservation network is performed on the newly generated virtual sample to recover the MixUp coefficient, where the semantic is preserved implicitly in the virtual sample. Meanwhile, an adversarial regularization is also conducted on the recovered mixup coefficient to ensure the authenticity of the virtual HSI. In such a way, the proposed AdvMixup sees the unseen HSI and knows the unknown semantic. Throughout the learning process, the feature extraction network is shared in both supervised and unsupervised branches to obtain a superior decision boundary. Experimental results on three HSI benchmark datasets demonstrate that AdvMixup achieves state-of-the-art performance in semi-supervised HSI classification. [ABSTRACT FROM AUTHOR]

Abstract: In this paper, a novel and effective pornographic image recognition method is proposed. Contributions of this paper include two aspects. (1) Due to the fact that the images are mostly stored and transmitted with JPEG compressed format on Internet, feature extraction is directly performed in the compressed domain. The exacted features include those derived from skin color regions, the results of image retrieval, human face and region of interest, as well as the global features of color and texture. (2) Data mining method is employed to search for the potential decision rules from large-scale image feature sets. Taken the misclassification cost and test cost into account, multi-cost sensitive decision tree is constructed first to improve the recognition speed and accuracy. Furthermore, the concept of pornography degree is introduced into the decision rules, which is output as the recognition results to represent the probability of the image being judged as pornographic. Experimental results show that, the recognition speed of the proposed method is almost three times faster than the classical pixel domain-based recognition method, and the recognition accuracy is also higher in terms of True Alarm Rate (TPR) and False Alarm Rate (FPR). [Copyright &y& Elsevier]

Highlights • A new two channel Convolutional Neural Network combining with Bi-directional Long Short-Term Memory achieves excellent performance for modulation recognition and protocol recognition. • Transfer learning is firstly introduced in waveform recognition filed to solve the model transferability problem across different recognition tasks, which can efficiently save the 60% training time and 50% training data. • A complete protocol signal dataset is provided for public research about waveform recognition in passive radar. Graphical abstract Abstract Passive radar capable of recognizing illumination of opportunities can improve the detection performance on account of its functional properties of environment adaptivity. Waveform recognition approaches based on Deep Learning can outperform traditional methods based on hand-crafted feature as shown in recent studies. In this paper, we propose a novel transferred deep learning waveform recognition method which makes use of multi-scale convolution and temporal dependency characteristics to improve the recognition performance. Firstly, we develop a two-channel convolutional neural networks combining with Bi-directional Long Short-Term Memory (TCNN-BL) architecture to extract features of different scales and merge past and future states. Then in order to solve the transferability problem across various sampling frequencies, we present a parameter transfer approach which initializes target domain classifier using source domain parameters. Based on our experiments on both public datasets and our own datasets, it can be demonstrated that the proposed approach significantly outperforms the state-of-the-art methods. [ABSTRACT FROM AUTHOR]

Perceptual image hashing and digital watermarking are two of the extensively investigated techniques for content authentication and copyright verification of digital images, respectively. One of the significant challenges of perceptual image hashing is a better tradeoff between robustness and discrimination. On the other hand, a major objective of digital watermarking techniques is to provide enhanced robustness and imperceptibility. In this paper, we present extensive experimental analysis, detailed laboratory forensic investigations, and application of the wave atom transform (WAT) based perceptual image hashing and digital watermarking techniques for printed images. Authenticity and copyright of printed images are verified via image hashing and digital watermarking techniques which are resilient against print-scan process distortions. To fulfill the requirement of a good balance between print-scan robustness and discrimination capabilities, we propose a compact, hybrid image hash using lower and middle frequency scale bands in the wave atom coefficients tiling. The hybrid image hash generation strategy is also advantageous due to its computational simplicity and reduced size of the hash value. Furthermore, we propose an enhanced multipurpose scheme which can support multiple applications, such as image authentication and copyright protection, simultaneously, for both digital-only and print-scanned images. Experimental analysis demonstrates the ability of wave atom based image hashing and watermarking schemes to provide authenticity and copyright verification for printed color images. [ABSTRACT FROM AUTHOR]

Keypoint detection is critical in image recognitions. Deep learning such as convolutional neural network (CNN) has recently demonstrated its tremendous success in detecting image keypoints over conventional image processing methodologies. The deep learning solutions, however, heavily rely on labeling target images for their reliability and accuracy. Unfortunately, most image datasets do not have all labels marked. To address this problem, this paper presents an effective and novel deep learning solution, Masked Loss Residual Convolutional Neural Network (ML-ResNet), to facial keypoint detection on the datasets that have missing target labels. The core of ML-ResNet is a masked loss objective function that ignores the error in predicting the missing target keypoints in the output layer of a CNN. To compensate for the loss induced by the masked loss objective function that likely results in overfitting, ML-ResNet is designed of a data augmentation strategy to increase the number of training data. The performance of ML-ResNet has been evaluated on the image dataset from Kaggle Facial Keypoints Detection competition, which consists of 7049 training images, but with only 2140 images that have full target keypoints labeled. In the experiments, ML-ResNet is compared to a pioneer literature CNN facial keypoint detection work. The experiment results clearly show that the proposed ML-ResNet is robust and advantageous in training CNNs on datasets with missing target values. ML-ResNet can improve the learning time by 30% during the training and the detection accuracy by eight times in facial keypoint detection. [ABSTRACT FROM AUTHOR]

In recent years, there have been increasing interests in addressing the issue of high-resolution inverse synthetic aperture radar (ISAR) imaging from sparse aperture (SA) data. The non-uniform rotation property of maneuvering target introduces non-stationary echo modulation to exhibit two-dimensional (2-D) migration through resolution cells (MTRC), which increases the difficulty of SA-ISAR imaging. In this paper, we focus on ISAR imaging and scaling of maneuvering target for effective MTRC correction and SA synthesis. Under a scaled non-uniform Fourier dictionary to include MTRC, SA maneuvering target imaging is formatted as sparse representation with maximum a posterior (MAP) estimation. Then, sparse ISAR imaging joint with parameter estimation is solved by using a two-step procedure. The estimated rational parameters are used to rescale the reconstructed ISAR image to extract the 2-D target geometry. Finally, the experiments based on simulated and measured data are performed to confirm the effectiveness of the proposed algorithm. [ABSTRACT FROM AUTHOR]

Comparing to the low-level local features, transform invariant low-rank textures (TILT) can in some sense globally rectify a large class of low-rank textures in 2D images, and thus more accurate and robust. However, the existing algorithms based on the alternating direction method (ADM) and the linearized alternating direction method with adaptive penalty (LADMAP), suffer from the weak robustness and the local minima, especially with plenty of corruptions and occlusions. In this paper, instead of exploiting optimization methods, we propose to build a hierarchical Bayesian model to TILT and then a variational method is implemented for Bayesian inference. Instead of point estimation, the proposed Bayesian approach introduces the uncertainty of the parameters, which has been proven to have much less local minima. Experimental results on both synthetic and real data indicate that our new algorithm outperforms the existing algorithms especially for the case with corruptions and occlusions. [ABSTRACT FROM AUTHOR]

This paper presents an approach of dual-band imagery based on heterogeneous sensor network which consists of infrared and optical sensors, and our approach of imaging reconstructs the three-dimensional shapes of the objects in some special environments such as smog, darkness, glare and so forth. Two dichroic mirrors are used for the construction of the binocular stereo vision system which consists of two optical (visible) and two infrared cameras. We make use of the complementarity between infrared and visible images to form two fused images as the source images of the stereo vision system. A stereo matching algorithm based on dual-band images is proposed as well. We use the gray, gradient direction and gradient magnitude of the pixels taken from the two bands of images to compute the disparity value of each pixel. Experiments show that the stereo matching algorithm is robust and not sensitive to the environmental variations. As a result, the dual-band stereo vision system has a good support for the target recognitions and measurements in different environments, such as all-weather traffic monitoring, searching and rescuing in fires, and so forth. [ABSTRACT FROM AUTHOR]

• Exhaustive survey of rotation invariant moments and transforms is presented. • Classification of the various rotation invariant techniques is discussed. • Analysis of various invariants is carried which is supported by experimental results. • Results are compared across various distance and training-based classifiers. • Recommendations for the selection of best invariant and classifier are given. The theory of moments and transforms is well established and widely applied to a number of computer vision, pattern recognition and image processing applications. A sub-class of these moments and transforms are used as rotation invariant global descriptors. There exist a number of techniques in literature whose authors claim superiority of their approaches. Consequently, many researchers face difficulties in deciding about the best or the near-best rotation invariant technique for their use. This paper reviews numerous state-of-the-art rotation invariant techniques based on moments and transforms by conducting a comparative performance analysis of various prominent approaches available in literature. To support the analysis of the survey, detailed experiments are conducted to compare performance of the various invariants using Zernike moments (ZMs) as a representative of all moments and transforms. The ZMs are selected because they are the most commonly used rotation invariant moments. Experiments are conducted on three standard databases which represent a wide variety of objects in different proportions of the shape and color cues. It is shown that the magnitude of moments of the gray-scale and multi-channel representation of color objects provides the overall best recognition results across a variety of classifiers. [ABSTRACT FROM AUTHOR]