U-Net is a widely used model for medical image segmentation, renowned for its strong feature extraction capabilities and U-shaped design, which incorporates skip connections to preserve critical information. However, its decoders exhibit information-specific preferences for the supplementary content provided by skip connections, instead of adhering to a strict one-to-one correspondence, which limits its flexibility across diverse tasks. To address this limitation, we propose the Task-Adaptive Mixture of Skip Connections (TA-MoSC) module, inspired by the Mixture of Experts (MoE) framework. TA-MoSC innovatively reinterprets skip connections as a task allocation problem, employing a routing mechanism to adaptively select expert combinations at different decoding stages. By introducing MoE, our approach enhances the sparsity of the model, and lightweight convolutional experts are shared across all skip connection stages, with a Balanced Expert Utilization (BEU) strategy ensuring that all experts are effectively trained, maintaining training balance and preserving computational efficiency. Our approach introduces minimal additional parameters to the original U-Net but significantly enhances its performance and stability. Experiments on GlaS, MoNuSeg, Synapse, and ISIC16 datasets demonstrate state-of-the-art accuracy and better generalization across diverse tasks. Moreover, while this work focuses on medical image segmentation, the proposed method can be seamlessly extended to other segmentation tasks, offering a flexible and efficient solution for diverse applications.
The rapid development of digital image inpainting technology is causing serious hidden danger to the security of multimedia information. In this paper, a deep network called frequency attention-based dual-stream network (FADS-Net) is proposed for locating the inpainting region. FADS-Net is established by a dual-stream encoder and an attention-based blue-associative decoder. The dual-stream encoder includes two feature extraction streams, the raw input stream (RIS) and the frequency recalibration stream (FRS). RIS directly captures feature maps from the raw input, while FRS performs feature extraction after recalibrating the input via learning in the frequency domain. In addition, a module based on dense connection is designed to ensure efficient extraction and full fusion of dual-stream features. The attention-based associative decoder consists of a main decoder and two branch decoders. The main decoder performs up-sampling and fine-tuning of fused features by using attention mechanisms and skip connections, and ultimately generates the predicted mask for the inpainted image. Then, two branch decoders are utilized to further supervise the training of two feature streams, ensuring that they both work effectively. A joint loss function is designed to supervise the training of the entire network and two feature extraction streams for ensuring optimal forensic performance. Extensive experimental results demonstrate that the proposed FADS-Net achieves superior localization accuracy and robustness on multiple datasets compared to the state-of-the-art inpainting forensics methods.
To address the performance degradation of violation action recognition models due to changing operational scenes in power grid operations, this paper proposes a Few-shot Adaptive Network (FSA-Net). The method incorporates few-shot learning into the network design by adding a parameter mapping layer to the classification network and developing a task-adaptive module to adjust the network parameters for changing scenes. A task-specific linear classifier is added after the backbone, allowing the adaptive generation of classifier weights based on the changing task scene to enhance the model’s generalizability. Additionally, the model uses a strategy of freezing the backbone network and iteratively updating only certain module parameters during training in order to minimize training costs. This approach addresses the challenge of iteratively updating difficulties in the original model, which are caused by limited image data following scene changes. In this paper, 2000 samples under power grid scenarios are used as the experimental dataset; the average recognition accuracy for violation actions is 81.77% for images after scene changes, which represents a 4.58% improvement when compared to the ResNet-50 classification network. Furthermore, the model’s training efficiency is enhanced by 40%. The experimental results show that the method enhances the performance of the violation action recognition model before and after scene changes and improves the efficiency of the iterative model by updating with a smaller sample size, lower model design cost, and lower training cost.
Olfactory stimuli exert a strong and direct influence on human emotions by rapidly activating limbic regions, making the olfactory pathway a compelling channel for emotion elicitation in multisensory multimedia applications. However, compared with the extensive research on emotions induced by visual and auditory stimuli, emotional responses elicited by olfactory stimuli have received little attention in affective computing. To facilitate research on olfactory-driven emotional dynamics, we construct a novel olfactory-evoked multimodal physiological dataset, comprising 74 participants across three emotion categories with synchronized electroencephalogram (EEG), electrocardiogram (ECG), and photoplethysmography (PPG) recordings. Building upon this dataset, we propose the Quality-aware Multimodal Expert Network (QMENet) for robust multimodal emotion recognition. QMENet assigns a dedicated expert to each modality and introduces a signal quality-aware router that dynamically modulates expert fusion weights according to modality reliability, effectively addressing signal quality variations in real world settings. To ensure stable and effective expert collaboration, we further design a two-stage optimization framework with expert orthogonality loss and the expert contribution calibration module, which alleviates expert homogenization and prevents excessive reliance on dominant modalities. Experimental results demonstrate that QMENet achieves a mean accuracy of 93.16% on the proposed dataset, outperforming all state-of-the-art baseline methods (p < 0.05), and exhibits strong generalization on a public multimodal emotion recognition benchmark.