YOLOv5s infrared small target detection model with improved feature pyramid and fused attention mechanism

With the development of deep learning, thermal infrared target detection has a wide range of applications in various fields. However, most existing deep learning models have high computational complexity, sensitivity to hyperparameters, low timeliness, and unsatisfactory results in practical applications. In this paper, we aim to realize the effect of small-sample, high-performance algorithms based on the lightweight target detection model YOLOv5s and a small-scale dataset, without expanding the data samples. In this paper, a YOLOv5s infrared small target detection model with improved feature pyramid and fused attention mechanism is proposed. By introducing a weighted bidirectional feature pyramid network (BiFPN) and a polarized self-attention mechanism (PSA), the model can quickly fuse multi-scale features and realize the fusion of pixel-level fine features into deep features, thus obtaining richer target information. Experimental results show that the improved algorithm improves the accuracy by 8.8% and the model is lightweight to meet the real-time requirements.