Lightweight Adaptive Perception Attention Model

With the in-depth research on deep learning-based object detection techniques, lightweight detection models have been widely applied in industrial domains. The current mainstream end-to-end detection models, such as the YOLO series, suffer from the problem of parameter sharing in traditional convolutional downsampling. In addition, the fusion of feature information across different scales through the Neck structure still has room for improvement in terms of attention to feature information at different scales. In this paper, we propose a method that replaces traditional convolutional downsampling with RFAConv to address the issue of parameter sharing in convolutional kernels. Furthermore, we introduce the PSA attention mechanism module between the model's Backbone and Neck to further focus on contextual information at different scales, generating superior pixel-level attention. Comparative experiments were conducted on the PASCAL VOC2012 dataset using different classical algorithms. The experimental results demonstrate that the YOLOv8n-RP network achieves higher detection accuracy.