This paper presents a new method of human activity recognition, which is based on R transform and dynamic time warping (DTW) after the key frame is extracted from a cycle. For a key binary human silhouette, R transform is employed to represent low-level features. The advantage of the R transform lies in its low computational complexity and geometric invariance. The DTW distance based on the extracted features are calculated and compared similarities to recognize activities. Compared with other methods, ours is superior because the descriptor is robust to frame loss in the video sequence, disjoint silhouettes and holes in the shape, and thus achieves better performance in similar activities recognition, simple representation, computational complexity and template generalization. Sufficient experiments have proved the efficiency.

Sparse representation has been proven to be a very effective technique for various image restoration applications. In this paper, an improved sparse representation based method is proposed to detect and estimate defocus blur of imaging sensors. Considering the fact that the patterns usually vary remarkably across different images or different patches in a single image, it is unstable and time-consuming for sparse representation over an over-complete dictionary. We propose an adaptive domain selection scheme to prelearn a set of compact dictionaries and adaptively select the optimal dictionary to each image patch. Then, with nonlocal structure similarity, the proposed method learns nonzero-mean coefficients’ distributions that are much more closer to the real ones. More accurate sparse coefficients can be obtained and further improve the performance of results. Experimental results validate that the proposed method outperforms existing defocus blur estimation approaches, both qualitatively and quantitatively.

The security and privacy of a system are urgent issues in achieving secure and efficient learning-based systems. Recent studies have shown that these systems are susceptible to subtle adversarial perturbations applied to inputs. Although these perturbations are difficult for humans to detect, they can easily mislead deep learning classifiers. Noise injection, as a defense mechanism, can offer a provable defense against adversarial attacks by reducing sensitivity to subtle input changes. However, these methods face issues of computational complexity and limited adaptability. We propose a multilayer filter defense model, drawing inspiration from filter-based image denoising techniques. This model inserts a filtering layer after the input layer and before the convolutional layer, and incorporates noise injection techniques during the training process. This model substantially enhances the resilience of image classification systems to adversarial attacks. We also investigated the impact of various filter combinations, filter area sizes, standard deviations, and filter layers on the effectiveness of defense. The experimental results indicate that, across the MNIST, CIFAR10, and CIFAR100 datasets, the multilayer filter defense model achieves the highest average accuracy when employing a double-layer Gaussian filter (filter area size of $3\times 3$, standard deviation of 1). We compared our method with two filter-based defense models, and the experimental results demonstrated that our method attained an average accuracy of 71.9%, effectively enhancing the robustness of the image recognition classifier against adversarial attacks. This method not only performs well on small-scale datasets but also exhibits robustness on large-scale datasets (miniImageNet) and modern models (EfficientNet and WideResNet).

Sparse representation is a powerful statistical technique that has been widely utilized in image restoration applications. In this paper, an improved sparse representation model regularized by a low-rank constraint is proposed for single image deblurring. The key motivation for the proposed model lies in the observation that natural images are full of self-repetitive structures and they can be represented by similar patterns. However, as input images contain noise, blur, and other visual artifacts, extracting nonlocal similarities only with patch clustering algorithms is insufficient. In this paper, we first propose an ensemble dictionary learning method to represent different similar patterns. Then, low-rank embedded regularization is directly imposed on inputs to regularize the desired solution space which favors natural and sharp structures. The proposed method can be optimized by alternatively solving nuclear norm minimization and $l_1$ norm minimization problems to achieve higher restoration quality. Experimental comparisons validate the superior results of the proposed method compared with other deblurring algorithms in terms of visual quality and quantitative metrics.

This paper presents a new classification method for single subject’s motion. We employ R transform descriptor and Linear-chain Conditional Random Fields for representation and classification. What it solves is that global features are described and adjacent states are independent. We extract binary silhouettes from a video sequence and segment them into groups by cycle after building the background model within Gaussian mixture model. Then low-level features are represented by R transform and principal vectors are obtained by Principal Component Analysis. After forming these vectors into a matrix, we utilize Linear-chain Conditional Random Fields to train and classify these groups, and demonstrate its usability in practice. Compared with the state-of-the-arts, our approach is simple in motion representation and independent between adjacent frames, and recognition accuracy improves to about 96% in average. It can be implemented in video surveillance where the background is known

Analysis of human activity and online anomaly detection from video sequences is one of the hottest and difficult research areas in computer visions. This paper describes a method for pedestrian gait classification in video sequence and deals with the classification of human gait types based on the notion that gait types can be analyzed into a series of consecutive postures types. First, silhouettes are extracted using the Background subtraction method which is combined with the time-stepping method. Then a method using recursion method for establishment of the standard gait state sequence is proposed. Mean-while, wavelet moment method is used to extract features of the human body image, and the result matrix leads to Discrete hidden Markov models. Finally, Discrete hidden Markov models is used for human posture training, modeling and activity matching to recognize the human activity. The experiment tests show some encouraging results also indicates the algorithm has very small leak-examining and mistake-examining-rate, also shows the capability of real-time performance, which indicate that the method could be a choice for solving the problem but more tests are required.