Anime Face Recognition Method Based on Multi-scale Channel Attention

In order to obtain more feature information of anime faces and improve the recognition rate of anime faces, this paper proposes a multi-scale attention anime face recognition model MCANet (Multi scale Channel Attention Network), which integrates multi-scale features and channel attention features to improve the accuracy of anime face recognition. Firstly, at the multi-scale level, cross channel attention XCA (Cross Channel Attention) and multi-scale feature extraction module MSM (Multi Scale Module) are proposed to enhance the multi-scale feature extraction capability. Secondly, at the level of channel attention, a multi-scale cross channel attention module MSXAM (Multi Scale Attention Cross Module) is proposed, which can extract multi-scale information and add attention to capture feature information. In addition, this paper also proposes the Focus Downsample Module (FDM), which uses the Focus operation to reduce information loss during the downsampling stage. The method proposed in this paper was tested on two publicly available anime face recognition datasets and compared with current mainstream deep learning methods. The accuracy of our method was superior to that of current mainstream methods, verifying the effectiveness of our method. In addition, multiple ablation experiments were conducted to further validate the effectiveness of the improved module proposed in this paper.