Object Detection Algorithm Based on Channel-Spatial Fusion Cascade Attention

Abstract

To solve the problem of low accuracy and high leakage rate of small object detection in a one-stage object detection model, a cascade attention mechanism (CAtt) is proposed. Cascading attention integrates channel attention and spital attention mechanisms that can be flexibly embedded into any object detection model, such as YoloV4, to improve the performance of the object detection model. The attention mechanism is perfected from two aspects: one is to integrate channel attention and spatial attention respectively so that the two are more complementary. Secondly, by introducing the skipping connection mechanism, the shallow detail information is connected with the deep semantic information, to solve the problem of information loss caused by the multiple convolutions of the model. Four typical attention mechanisms were selected to compare public datasets such as PASCAL VOC2007 and PASCAL VOC2012. A large number of experimental results show that this method significantly improves the performance of the object detection model, especially for the PASCAL VOC2007. Embedding the method in YoloV4 resulted in a 4.77% increase in the mAP index of the original model.