Fusion of triplet attention mechanism and dynamic regularization in YOLO for aerial infrared small target detection

In the application of aerial infrared small object detection, images are typically characterized by low contrast and complex backgrounds. As a result, the feature maps extracted by traditional detection methods have a relatively low resolution, leading to suboptimal detection performance. To address this issue, YOLO-EFT was designed specifically based on aerial infrared target datasets. Firstly, the model effectively reduces the number of parameters by introducing a convolutional structure with a progressive training strategy. Secondly, by adding a feature fusion layer focal block in the shallow features, it can capture multi-scale feature interactions and enrich the feature representation for small target detection. Finally, a triplet attention mechanism is added to the feature extraction network, enabling the model to more comprehensively capture the key features among various dimensions of sequence data. Through testing and verification on the public dataset, compared to YOLOv5s, the mAP of YOLO-EFT increased by 12.2% and the network parameter quantity was reduced by 40.7%, effectively reflecting the improvement in the detection performance of aerial infrared small target detection.