MiniYOIR: ultra-lightweight infrared remote sensing image object detection algorithm based on YOLOv8

Object detection in infrared remote sensing images has been widely applied in daily life. However, due to the limited computing power of remote sensing equipment, only lightweight target detection algorithms can be used, which often suffer from accuracy loss. Therefore, we propose an ultra-lightweight target detection algorithm named MiniYOIR to address the above issues. Specifically, we utilize an improved reversible column networks for multi-scale feature extraction to adapt to low-resolution infrared images. Meanwhile, we employ depthwise separable convolution to reduce the model parameters and introduce the fusion block module in the neck to enhance the target detection capability for infrared images. Furthermore, we propose an efficient multi-scale group convolution module, which processes feature maps from different channels to enhance feature fusion capabilities. Finally, we optimize the model’s Intersection over Union using inner Intersection over Union, effectively improving detection accuracy. Experimental results on the infrared remote sensing dataset HIT-UAV show that MiniYOIR achieves a 1% improvement in mAP0.5 compared with YOLOv8n. On the AU-AIR dataset, the mAP0.5 increases by 3%. Simultaneously, the number of parameters is reduced by 40%, and the GFLOPs are reduced by 35%, making it deployable on small devices and satisfying real-time detection requirements.