Infrared small target detection based on YOLO v4

Abstract

Infrared imaging is highly stealthy, reconnaissance and resistant to interference and currently has numerous practical applications in many fields, particularly in the military and civilian sectors. Infrared images are prone to drowning small targets in the background due to the lack of textural detail and the presence of a large amount of background noise. Small target detection on infrared images is therefore very challenging, and traditional model-driven algorithms are no longer very applicable in the face of complex and variable infrared images, so more and more people are turning to deep learning methods. We have designed a target detection network based on YOLO v4 with the following improvements, using GhostNet as the backbone network used to extract features and achieving a lighter weighting of the backbone network. We added a coordinate attention mechanism to the design of the feature aggregation process of the network, so that the network senses the spatial location relationship while the channel is attended to, enhancing the network’s ability to localise targets. To further reduce the number of network parameters, by using depthwise separable convolution in the feature aggregation part and detection part instead of normal convolution. Our experimental comparisons on the dataset NUAA-SISRT show that Ghost-CA-YOLO v4 is the leader in terms of detection accuracy and parameter efficiency. The model has an mAP metric of 73.31% and an F1 value of 0.78. The number of parameters is 44.8M, which is 1/5 of that of YOLO v4.