A yolov8-based lightweight detection model for different perspectives infrared images

In recent years, the development of edge devices has enabled the capture and detection of infrared images from different perspectives. However, infrared image detection from different perspectives often results in low accuracy, massive model parameters, and high computing burden. To address these challenges, we propose an enhanced lightweight infrared target detection method leveraging YOLOv8n. Our method adopts the ADown downsampling operation in the model’s backbone, which gradually reduces the spatial size of the feature map while increasing its depth. This operation enables the network to capture and process information at various scales more effectively, enhancing its feature representation capabilities. Additionally, our method incorporates the Triplet Attention mechanism to enhance the effectiveness of feature extraction. Finally, to optimize the feature pyramid and path aggregation network, we propose the ASF-DynamicScalSeq structure, utilizing GSConv, VoVGSCSP to balance accuracy and processing speed. The ASF-ScalSeg optimizes spatial-scale features, while the DySample dynamic upsampling function enhances performance without compromising efficiency. In infrared datasets, including those from UAV and vehicle perspectives, our algorithm shows a 2.9% and 1% increase in mAP50, respectively, compared to YOLOv8n. Additionally, our approach reduces parameters by 21.9%, decreases computational burden by 13.4%, and enhances inference speed by 3.9% and 7.8%, respectively. To validate the effectiveness and robustness of our method, we conduct experiments on FLIR and aerial infrared datasets. The results show a 0.9% and 1% increase in mAP50, a 21.9% reduction in parameters, a 13.4% decrease in computational burden, and a 5.9% and 7.8% improvement in inference speed, respectively, compared to YOLOv8n.