Search

With the rise of short video and the development of self-media, video has become a type of frequently used data, thus the protection of video data is particularly important. Digital watermarking is a very effective digital authentication and copyright protection technology that hides a piece of data called watermark into digital multimedia content. However, most existing video watermark algorithms fail to deal with the videos under geometric distortions, and they only focus on copyright protection, ignoring the task of tamper detection. This paper proposes a multipurpose video watermarking algorithm that is resistant to rotation and scaling attacks while providing both copyright protection and tamper detection. We design a DFT template embedding method to make the watermarking algorithm resistant to rotation and scaling attacks efficiently. This method can detect the rotation and scaling that the video has undergone during the watermark extraction process, and then extract the copyright watermark information based on geometric correction. In addition, we propose a robust watermark embedding method with tamper detection function, which embeds copyright watermark information with validation bits in the DCT coefficients of I frames. By embedding the copyright watermark block by block, the algorithm can not only improve the robustness, but also enable the system to detect tampering. Experiments show that the proposed algorithm has good robustness to various attacks such as compression, filtering, rotation, scaling, frame rate changing and other inter-frame operations, and it outperforms previous geometrically robust algorithm by an average of 12.8% under geometric attacks on the normalized correlation values. Also, we show that our algorithm is effective in tamper detection by localizing the distorted areas in the video frames. [ABSTRACT FROM AUTHOR]

In this paper, we focus on how to better represent video data in the field of action recognition. We propose a new image feature named phase spectrum reconstruction map to facilitate action recognition, which extracts the contour features of RGB frame images from a video clip beneficial for action recognition. We demonstrate the effectiveness of such a feature with ablation experiments using the channel-based feature fusion method and two-stream method. Also, we verify that the reconstructed map does contain motion-related features and can be learned by convolutional neural networks only using the reconstructed map as input features. Our method is trained and evaluated using the benchmark datasets HMDB-51 and UCF-101, and both show significant improvements over other methods without adding the reconstructed map features. [ABSTRACT FROM AUTHOR]

The major obstacle in semantic segmentation is that it requires a large number of pixel-level labeled data to train an effective model. In order to reduce the cost of annotation, weakly-supervised methods use weaker labels to overcome the need for per-pixel labels, while zero-shot methods transfer the knowledge learned from seen classes to unseen classes to reduce the number of classes that need to be labeled. To further alleviate the burden of annotation, we introduce a more challenging task of Weakly-supervised Zero-shot Semantic Segmentation (WZSS): learning models which only utilize image-level annotation of seen classes to segment images containing unseen objects. To this end, we propose a Dual Semantic-Guided (DSG) model which is double guided by semantic embeddings of classes to obtain classification scores and localization maps. By ignoring the localization maps with low classification scores, our proposed framework can generate prediction segmentation masks. To improve our model's performance, we propose a simple stochastic selection on semantic embeddings during inference, which explores the difference between image-level class embeddings and pixel-level class embeddings. This simple approach increases our model's performance in terms of hIoU from 25.9 to 31.8. In addition, compared with some zero-shot semantic segmentation methods, our method delivers better results in terms of hIoU (31.8) and mIoU u (22.0) on the PASCAL VOC 2012 dataset with less supervision information. [ABSTRACT FROM AUTHOR]

The development of the Internet, together with the progress of multimedia processing techniques, has led to the problems of data piracy, data tampering and illegal dissemination. Digital watermarking is an effective approach to data authentication and copyright protection. This paper proposes a geometrically robust multi-bit video watermarking algorithm based on 2-D DFT (two-dimensional discrete Fourier transform). While most of the existing video watermarking schemes require synchronization to extract the watermark from rotated or scaled videos, which is time-consuming and affects the accuracy, the proposed method can do direct extraction without performing synchronization for videos attacked by rotation, scaling or cropping. For embedding the watermark, circular templates in DFT domain are transformed into spatial masks and added to the video frames in spatial domain. A perceptual model based on local contrast is applied to keep the fidelity of the watermarked video. We also propose an accurate and efficient extraction method which is based on the cross-correlation between the Wiener-filtered DFT magnitude and the stretched template sequence in polar coordinates. Experimental results show that the proposed algorithm is robust against various kinds of attacks, such as compression, filtering, rotation, scaling, cropping, frame averaging and frame rate changing. [ABSTRACT FROM AUTHOR]

Action recognition plays a central role in intelligent surveillance system, game-control, human-computer interaction, and so on. In this work, we design a multi-task framework that improves the recent Spatial-Temporal Graph Convolutional Networks (ST-GCN) for skeleton-based action recognition by introducing the attention mechanism and co-occurrence feature learning. Specifically, we use an attentional branch to pay more attention to more discriminating features and aggregates co-occurrence features from all joints globally in another branch. Additionally, our multi-task framework exploits the inherent correlation between branches to further enhance the classification accuracy and convergence speed. Experiments have been carried out on NTURGB+D and Kinetics human action dataset. The results clearly show that the accuracy of the proposed multi-task framework are distinguishably higher than ST-GCN and other mainstream methods for 3D action recognition. [ABSTRACT FROM AUTHOR]

The robust watermarking scheme has very broad application scenarios. For robust watermarking methods, the transform domain based watermarking algorithms are popular. Indeed, these algorithms have strong robustness, but the visual effects may be affected. In order to balance the robustness and imperceptibility, a histogram based watermarking scheme is proposed. In this scheme, in order to enhance the robustness and imperceptibility, the low-frequency construction method (LFCM) is proposed. In LFCM, the most suitable pixels for modifications are selected by calculating the weights of pixels at different positions, and these pixels will be modified in the histogram based embedding process. After the modification, the low-frequency of the image will enhance, and the robustness of the watermarking will enhance at the same time. Experiment results show that the proposed scheme can resist compression, scaling, noise, and median blur attacks, and can resist rotation attacks to a certain extent. The proposed scheme is a non-blind watermarking scheme, and non-blind watermarking scheme is feasible when this scheme has good performance. [ABSTRACT FROM AUTHOR]

During the image acquisition process of face recognition, the obtained face images are affected inevitably by varied illumination and position in different environment. Local Binary Pattern (LBP) operator is used to decrease illumination effectiveness. Improved Pairwise-constrained Multiple Metric Learning method (IPMML) is proposed as a classification metric in our prior work, which solves the misalignment problem in a better way compared with PMML. To solve the high computation complexity of IPMML, Linear Discriminant Analysis (LDA) is performed before IPMML. Thus, a face recognition method based on LBP and IPMML is proposed, which can overcome the illumination and misalignment problems. LBP is selected to extract texture features of face images firstly. Second, LDA is applied to reduce the dimension. Then the fisher features are divided into sub-blocks according to the dimension of features and every block is a column vector. Fourth, a classification metric -- IPMML is used to obtain the optimum Mahalanobis matrix. Fifth, the Mahalanobis matrix is used to compute the final discriminative distance. Finally, the Nearest Neighborhood Classifier (NNC) is applied to classify face images. The experimental results show that the proposed method can achieve high recognition rates and is robust to illumination and facial expression variation, especially for misaligned face images. [ABSTRACT FROM AUTHOR]

In this paper, we propose a novel face feature extraction approach based on Local Binary Pattern (LBP) and Two Dimensional Locality Preserving Projections (2DLPP) to enhance the texture features and preserve the space structure properties of a face image. LBP is firstly used to remove the effect of illumination and noise, which would enhance the detailed texture characteristics of face images. Then 2DLPP is performed to extract some prominent features and decrease the image dimension with space structure information. The Nearest Neighborhood Classifier (NNC) is used to recognize a face image at the end. In addition, the rule for dimension selection is studied from the results of experiments about choosing an appropriate feature dimension by 2DLPP computation. The experimental results on the Yale, the extended Yale B and CMU PIE C09 benchmark datasets showed that the proposed face feature extraction and recognition method achieves a better performance in comparison with similar techniques, and the proposed dimension selection rule can give an appropriate feature dimension in 2DLPP. [ABSTRACT FROM AUTHOR]

Frame insertion, deletion and duplication are common inter-frame tampering operations in digital videos. In this paper, based on similarity analysis, a passive-blind forensics scheme for video shots is proposed to detect inter-frame forgeries. This method is composed of two parts: HSV (Hue-Saturation-Value) color histogram comparison and SURF (Speeded Up Robust Features) feature extraction together with FLANN (Fast Library for Approximate Nearest Neighbors) matching for double-checking. We mainly calculate H-S and S-V color histograms of every frame in a video shot and compare the similarity between histograms to detect and locate tampered frames in the shot. Then we utilize SURF feature extraction and FLANN matching to further confirm the forgery types in the tampered locations. Experimental results demonstrate that the proposed detection method is efficient and accurate in terms of forgery identification and localization. In contrast to other inter-frame forgery detection methods, our scheme can detect three kinds of forgery operations and has its own superiority and applicability as a passive-blind detection method. [ABSTRACT FROM AUTHOR]

In this paper, a new blind watermarking for copyright protection of color images based on the U matrix through Quaternion Singular Value Decomposition (QSVD) is proposed. The proposed method represents the color image with a quaternion matrix, so that it can deal with the multichannel information in a holistic way. Then the array of pure quaternion is divided into non-overlapping blocks and we perform QSVD on each block to get its U matrix. The watermark is inserted into the optimally selected coefficients of the quaternion elements in the first column of the U matrix. Besides, in the procedures of watermark insertion and extraction, ensuring higher fidelity and robustness to several possible image attacks have been considered. The experimental results show that the proposed method outperforms existing schemes in terms of robustness and invisibility. [ABSTRACT FROM AUTHOR]

Vector quantization (VQ) has been successfully used in data compression and feature extraction areas. Codebook design is the essential step of VQ. The K-means algorithm is a famous data clustering technique which is also an efficient codebook design scheme. The main disadvantages of K-means algorithm lie in that the initial cluster centroids greatly affect the convergence speed and the final clustering performance. In the past two decades, many codebook initialization techniques have been proposed. However, most of these techniques do not make full use of the features of the training vectors, and some techniques require high extra computational load. This paper presents an efficient and simple technique for the conventional K-means algorithm based on feature classification and grouping. Firstly, all training vectors are classified into sixteen categories based on a two-level classifier including an edge classifier and a contrast classifier. Then, the training vectors in each category are sorted based on their norm values and divided into groups. Each group has the same size, and the centroid vector of each group is calculated as an initial codeword. Experimental results show that, compared with several typical initialization techniques, our technique can obtain a better codebook along with a faster convergence speed in a shorter time. [ABSTRACT FROM AUTHOR]