A target detection model based on parallel interactive feature extraction and attention fusion structure

Abstract

In this paper, an anomaly inspection algorithm is designed with the infrared anomaly inspection task of substation. Firstly, a parallel interaction feature extraction structure is designed, which can strengthen the information interaction ability of the main part of the model, and the P and R indexes of the model are improved by 4% and 8.7%, respectively; secondly, a parallel encoder structure is designed to strengthen the information interaction between the higher levels of the network, and strengthen the ability of recognizing the small targets; thirdly, a feature attention fusion structure is designed, which can pay attention to the regions to be detected other than the background and the important information channels, so that the information extracted from the features is fully utilized and the model training accuracy is improved. Finally, based on the above innovations, this paper finally proposes a target detection model for substation infrared anomaly inspection, and the mAP50 on substation infrared inspection dataset, FLIRv2, VisDrone2019, and VOC dataset reaches 98.9%, 49.7%, 40.2%, and 79.4%, respectively, which proves the validity of the model design in this paper.