Referring video object segmentation (R-VOS) is a fundamental vision-language task which aims to segment the target referred by language expression in all video frames. Existing query-based R-VOS methods have conducted in-depth exploration of the interaction and alignment between visual and linguistic features but fail to transfer the information of the two modalities to the query vector with balanced intensities. Furthermore, most of the traditional approaches suffer from severe information loss in the process of multi-scale feature fusion, resulting in inaccurate segmentation. In this paper, we propose DCT, an end-to-end decoupled cross-modal transformer for referring video object segmentation, to better utilize multi-modal and multi-scale information. Specifically, we first design a Language-Guided Visual Enhancement Module (LGVE) to transmit discriminative linguistic information to visual features of all levels, performing an initial filtering of irrelevant background regions. Then, we propose a decoupled transformer decoder, using a set of object queries to gather entity-related information from both visual and linguistic features independently, mitigating the attention bias caused by feature size differences. Finally, the Cross-layer Feature Pyramid Network (CFPN) is introduced to preserve more visual details by establishing direct cross-layer communication. Extensive experiments have been carried out on A2D-Sentences, JHMDB-Sentences and Ref-Youtube-VOS. The results show that DCT achieves competitive segmentation accuracy compared with the state-of-the-art methods.

The rapid development of deepfake facial technology has led to facial fraud, posing a significant threat to social security. With the advent of diffusion models, the realism of forged facial images has increased, making detection increasingly challenging. However, the existing detection methods primarily focus on identifying facial forgeries generated by generative adversarial networks; they may struggle to generalize when faced with novel forgery techniques like diffusion models. To address this challenge, a multi-branch network with multi-feature enhancement (M2EH) model for improving the generalization of facial forgery detection is proposed in this paper. First, a multi-branch network is constructed, wherein diverse features are extracted through the three parallel branches of the network, allowing for extensive analysis into the subtle traces of facial forgeries. Then, an adaptive feature concatenation mechanism is proposed to integrate the diverse features extracted from the three branches, obtaining the effective fused representation by optimizing the weights of each feature channel. To further enhance the facial forgery detection ability, spatial pyramid pooling is introduced into the classifier to augment the fused features. Finally, independent loss functions are designed for each branch to ensure the effective learning of specific features while promoting collaborative optimization of the model through the overall loss function. Additionally, to improve model adaptability, a large-scale deepfake facial dataset, HybridGenFace, is built, which includes counterfeit images generated by both generative adversarial networks and diffusion models, addressing the limitations of existing datasets concerning a single forgery type. Experimental results show that M2EH outperforms most of the existing methods on various deepfake datasets.