Low-resolution Infrared Image Target Detection based on Improved YOLOv5

In the context of border defense and peacekeeping operations carried out in challenging environments, conventional visible light target detection methods prove to be poor in performance. Therefore, the utilization of infrared images detection is predominantly employed. In this study, we focus on the improvement of target detection performance of YOLOv5 model in infrared images. To enhance the original YOLOv5 network structure, we draw inspiration from BoTNet and introduce the BoT3 block and Multi Head Self Attention mechanism. The convolutional blocks in the neck were substituted with GSConv modules. Additionally, the Focal Loss was replaced with VariFocal Loss to achieve a more balanced weighting of samples. The experimental results display that the enhanced YOLOv5 model outperforms the original model and achieve better detection outcomes when applied to low-resolution infrared images.