Rapider-YOLOv7: Lightweight Infrared Object Detection Network

YOLOv7-tiny, as a lightweight variant of YOLOv7, boasts advantages of fast runtime and fewer parameters. However, when directly applied to infrared object detection, YOLOv7-tiny still faces challenges such as weak extraction of fine details, loss of semantic information, and high resource consumption. To address these issues, we propose a comprehensive balance between detection speed and accuracy in the Rapider-YOLOv7 object detection network. Firstly, it introduces the Exponential Spatial Pyramid Pooling with Channel Spatial Pyramid (Soft-SPPCSP) module to enhance the capability of the original model in extracting low contrast information from infrared images. Secondly, it designs the FO-CA attention module to strengthen the model's feature extraction capability for infrared images with weak textures. Furthermore, it devises the lightweight SFaster-net network, which incorporates SPDconv convolution for feature extraction to accelerate the model's feature extraction speed and enhance its capability to extract features from low-resolution images. Experimental results on the FLIR-v2 infrared dataset demonstrate that the mAP of the Rapider-YOLOv7 model reaches 62.6%, which is 2.9% higher than that of the original YOLOv7-tiny. Moreover, the FPS value on the RTX4060ti reaches 93.