Video emotion recognition is widely used due to its alignment with the temporal characteristics of human emotional expression, but existing models have significant shortcomings. On the one hand, Transformer multi-head self-attention modeling of global temporal dependency has problems of high computational overhead and feature similarity. On the other hand, fixed-size convolution kernels are often used, which have weak perception ability for emotional regions of different scales. Therefore, this paper proposes a video emotion recognition model that combines multi-scale region-aware convolution with temporal interactive sampling. In terms of space, multi-branch large-kernel stripe convolution is used to perceive emotional region features at different scales, and attention weights are generated for each scale feature. In terms of time, multi-layer odd-even down-sampling is performed on the time series, and odd-even sub-sequence interaction is performed to solve the problem of feature similarity, while reducing computational costs due to the linear relationship between sampling and convolution overhead. This paper was tested on CMU-MOSI, CMU-MOSEI, and Hume Reaction. The Acc-2 reached 83.4%, 85.2%, and 81.2%, respectively. The experimental results show that the model can significantly improve the accuracy of emotion recognition.

Discriminative region localization and efficient feature encoding are crucial for fine-grained object recognition. However, existing data augmentation methods struggle to accurately locate discriminative regions in complex backgrounds, small target objects, and limited training data, leading to poor recognition. Fine-grained images exhibit “small inter-class differences,” and while second-order feature encoding enhances discrimination, it often requires dual Convolutional Neural Networks (CNN), increasing training time and complexity. This study proposes a model integrating discriminative region localization and efficient second-order feature encoding. By ranking feature map channels via a fully connected layer, it selects high-importance channels to generate an enhanced map, accurately locating discriminative regions. Cropping and erasing augmentations further refine recognition. To improve efficiency, a novel second-order feature encoding module generates an attention map from the fourth convolutional group of Residual Network 50 layers (ResNet-50) and multiplies it with features from the fifth group, producing second-order features while reducing dimensionality and training time. Experiments on Caltech-University of California, San Diego Birds-200-2011 (CUB-200-2011), Stanford Car, and Fine-Grained Visual Classification of Aircraft (FGVC Aircraft) datasets show state-of-the-art accuracy of 88.9%, 94.7%, and 93.3%, respectively.

Ship re-identification (Re-ID) aims to match ship identities across disjoint camera views and separated time periods, which is critical for maritime target tracking and law enforcement. In real-world surveillance, variations in target distance and viewing angle frequently produce partial views and occlusions, leading to missing geometric components and fragmented appearance cues. Such incomplete observations substantially degrade the robustness and generalization of conventional single-frame methods that rely on global appearance representations. To address these challenges, this study proposes a new ship re-identification framework based on dual-stream feature decoupling and temporal variational Bayesian inference. The proposed method explicitly disentangles ship representations into appearance and structural streams, and leverages multi-frame temporal context to infer missing components and enhance discriminability under partial visibility. Specifically, a ResNet-based splitter trained adversarially against two discriminators is employed to decouple the input representation into separate feature streams. The decoupled streams are then modeled over time using a bidirectional LSTM (BiLSTM) together with a visibility-probability estimator. A graph-structured spatial prior, parameterized via a graph attention network (GAT), serves as the variational prior. Given sequential observations, the variational inference module estimates posterior distributions for missing components and performs probabilistic completion in the latent space. The framework is trained end-to-end using cross-entropy and triplet losses. Extensive experiments on the Ship-CH dataset demonstrate that our method achieves 85.67% mAP and 93.67% Rank-1 accuracy, exhibiting superior robustness under occlusion and partial visibility.

Robust watermarking requires finding invariant features under multiple attacks to ensure correct extraction. Deep learning has extremely powerful in extracting features, and watermarking algorithms based on deep learning have attracted widespread attention. Most existing methods use $3\times 3$ small kernel convolution to extract image features and embed the watermarking. However, the effective perception fields for small kernel convolution are extremely confined, so the pixels that each watermarking can affect are restricted, thus limiting the performance of the watermarking. To address these problems, we propose a watermarking network based on large kernel convolution and adaptive weight assignment for loss functions. It uses large-kernel depth-wise convolution to extract features for learning large-scale image information and subsequently projects the watermarking into a high-dimensional space by $1\times 1$ convolution to achieve adaptability in the channel dimension. Subsequently, the modification of the embedded watermarking on the cover image is extended to more pixels. Because the magnitude and convergence rates of each loss function are different, an adaptive loss weight assignment strategy is proposed to make the weights participate in the network training together and adjust the weight dynamically. Further, a high-frequency wavelet loss is proposed, by which the watermarking is restricted to only the low-frequency wavelet sub-bands, thereby enhancing the robustness of watermarking against image compression. The experimental results show that the peak signal-to-noise ratio (PSNR) of the encoded image reaches 40.12, the structural similarity (SSIM) reaches 0.9721, and the watermarking has good robustness against various types of noise.