21 January 2025 in Applied Intelligence
1 September 2023 in International Journal of Digital Earth
14 January 2023 in Neural Networks
11 October 2022 in IEEE Transactions on Biometrics, Behavior, and Identity Science
Gait recognition refers to video-based biometric techniques for recognizing subjects by walking patterns under a long-distance situation. Despite the progress of existing gait recognition methods, the recognition ability of subjects in carrying situations is still limited (e.g., carrying bags or wearing coats or jackets). To address this issue, this paper proposes to extract human gait information from different ranges of receptive fields, providing richer internal features of the deep network. Moreover, we propose two attention mechanisms, Local Pyramid Attention and Global Attention Fusion Learning, to focus on the key features in human gait from different perspectives. Depending on the different attention mechanisms employed in the network, three network variants are derived, where the Gait Pyramid Attention Network (GPAN) contains two attention mechanisms, while GPAN-P and GPAN-L contain a single attention mechanism. We evaluated our method on two large datasets, CASIA-B and OUMVLP. Experiments show that the proposed network gives an average rank-1 accuracy of 97.8% on CASIA-B under normal walking conditions. We also achieve 94.2% and 81.8% accuracy on CASIA-B dataset under the complex bag-walking and coat-walking scenarios, which are dramatically superior to the state-of-the-art methods.
19 January 2024 in IEEE Transactions on Biometrics, Behavior, and Identity Science
Face super-resolution (SR) algorithms have recently made significant progress. However, most existing methods prefer to employ texture and structure information together to promote the generation of high-resolution features, neglecting the mutual encouragement between them, as well as the effective unification of their own low-level and high-level information, thus yielding unsatisfactory results. To address these problems, we propose an implicit mutual learning of dual-branch networks for face super-resolution, which adequately considers both extraction and aggregation of structure and texture information. The proposed approach consists of four essential blocks. First, the deep feature extractor is equipped with a deep feature reinforcement module (DFRM) based on two-stage cross-dimensional attention (TCA), which behaves in the texture enhancement and structure reconstruction branches, respectively. Then, we elaborate two information exchange blocks for two branches, one for the first information exchange block (FIEB) from the texture branch to the structure branch and one for the second information exchange block (SIEB) from the structure branch to the texture branch. These two interaction blocks perform further fusion enhancement of potential features. Finally, a hybrid fusion network (HFNet) based on supervised attention executes adaptive aggregation of the enhanced texture and structure maps. Additionally, we use a joint loss function that modifies the recovery of structure information, diminishes the use of potentially erroneous information, and encourages the generation of realistic face images. Experiments on public datasets show that our method consistently achieves better quantitative and qualitative results than SOTA methods.
19 January 2022 in IEEE Signal Processing Letters
Video action recognition task involves modeling spatiotemporal information, and efficiency is critical to capture spatiotemporal dependencies in the video. Most existing models rely on optical flow information to capture the dynamic visual tempos between consecutive video frames. Although impressive performance can be achieved by combining optical flow with RGB, the time-consuming nature of optical flow computation cannot be ignored. Moreover, 3D CNN has successfully modeled spatiotemporal information, yet the enormous computational volume is unsuitable for real-time action recognition. In this letter, we propose a novel lightweight video feature extraction strategy that achieves better recognition performance with lower FLOPs. In particular, we perform convolution on the video cube from three orthogonal angles to learn its appearance and motion features. Compared with the computational volume of 3D CNN, our proposed method is more economical and thus meets the lightweight requirements. Extensive experimental results on public Something Something-V1 $\&$ V2 and Diving48 datasets show our approach achieves the state-of-the-art performance.