An infrared remote sensing image target detection algorithm based on an improved YOLOv8

Infrared remote sensing images possess significant application value in military reconnaissance and disaster monitoring due to their all-weather imaging capabilities. Nonetheless, variables such as low image resolution, intricate backgrounds, and a low signal-to-noise ratio render small target detection challenging, frequently resulting in missed detections. This work provides an improved technique based on YOLOv8n to augment the detection performance of tiny targets in infrared remote sensing images. The method incorporates a spatial-to-depth conversion module (SPD-Conv) that minimizes detail loss during downsampling by restructuring spatial information into channel information. It also integrates cavity convolution techniques to broaden the receptive field and improve the acquisition of cross-layer semantic features. The C2f-Global Residual Coordinate (GRC) module, which integrates channel, spatial, and coordinate attention mechanisms, is incorporated into the C2f module to enhance the representation of target boundary and positional information. Simultaneously, the dynamic non-monotonic focusing WIoU_v3 loss function is employed to mitigate the adverse effects of low-quality samples on model training and enhance the precision of bounding box localization. The experimental results indicate that, under comparable parameter counts and inference speeds, the enhanced model exhibits superior performance across multiple metrics: precision has risen to 83.9%, recall has increased by 18.6 percentage points, mAP50 has improved from 72.4% to 90.5%, and mAP50-95 has increased by 18.1 percentage points. The results confirm the efficacy and practical significance of the suggested strategy in intricate infrared remote sensing situations.