基于LoRA微调与小波注意力增强的CLIP表情识别方法OA

In facial expression recognition tasks,the scarcity of high-quality annotated data—particularly for minority categories such as disgust and fear—poses a significant constraint on model performance.To address the data insufficiency of specific expression categories,this paper proposes a facial expression recognition method based on the CLIP vision large model,aiming to enhance the representational capacity for scarce expressions via a parameter-efficient adaptation mechanism.Specifically,an adaptive adjustment strategy is introduced for the CLIP visual encoder,with the base model parameters kept frozen during LoRA fine-tuning to achieve effective alignment between pre-trained visual representations and the facial expression recognition task.Furthermore,this paper presents a wavelet channel attention enhancement module,which suppresses high-frequency random noise in features through multi-scale decomposition and channel-adaptive weighting mechanisms,thereby strengthening the capture of subtle expression dynamics.Finally,a gated fusion module is proposed,which employs a learnable channel-level weight allocation mechanism to achieve effective complementarity between local detail features and high-level semantic features.Extensive experimental results demonstrate that the proposed method achieves high recognition accuracy,validating the effectiveness and robustness in expression recognition tasks.