Lightweight Infrared Target Detection Algorithm based on Adaptive Feature Enhancement

Abstract

A lightweight infrared target detection algorithm AFE-YOLO with adaptive feature enhancement is proposed to address the problems of blurred infrared target features and low detection accuracy due to the difficulty of distinguishing them from the background in complex backgrounds. to construct the backbone network, which effectively enhances the target features and improves the feature extraction capability of the model while keeping the model lightweight; finally, in the feature fusion stage, channel blending and adaptive feature enhancement module are introduced to further adjust the fused features at different levels. Experiments are conducted on the public dataset FLIR, and the results show that the number of parameters and the computation amount of AFE-YOLO are decreased by 21.26% and 7.15%, respectively, compared with YOLOv5n, while the accuracy is improved by 2.1%. In addition, compared with other lightweight models, the algorithm in this paper still maintains the balance of lightweight and accuracy.