An object detection method based on multilayer attention mechanism

Object detection faces the challenge of multi-scale target recognition and insufficient feature expression in complex scenes. This study proposes a visual entity localization technique built upon a multi-layer attention mechanism. Faster R-CNN combined with Feature Pyramid Network (FPN) is used as the benchmark framework, and multi-scale feature fusion is optimized into Faster R-CNN-FPN (FR-CNN-FPN) algorithm model by introducing hierarchical attention module. The method embeds channel-space co-attention mechanism in the lateral connection path of the FPN;The detection accuracy (AP) of this method reaches 95.76%, which is higher than Faster R-CNN, and the detection performance of small targets (APs) and large targets (APm) is reached by 42.6% and 70.2% respectively. In addition, through the lightweight attention head design, the model computational overhead is only reached by 90.2%, and the inference speed is maintained at 36 FPS. The experimental results verify the effectiveness of multi-layer attention mechanism in multi-scale feature optimization, and provide a high-precision solution for object detection in complex scenes.