Small Target Detection in UAV Aerial Images Based on Sliding Window and Improved Yolov8

Target detection in drone aerial images is widely used in agriculture, urban management, military defense and other fields. Aerial images have problems such as small targets, dense targets and complex backgrounds, which bring difficulties and challenges to target detection. This paper uses YOLOv8n as the benchmark model to realize target detection in drone aerial images. First, each image is divided into multiple overlapping small areas through a sliding window, and the labels are regenerated, all areas are trained and predicted, and finally the prediction results are restored to the original image; secondly, a lightweight attention module (CBAM) and a bidirectional feature pyramid network (BiFPN) are added to YOLOv8. CBAM introduces channel attention and spatial attention mechanisms, which can enhance the model's attention to important features. BiFPN introduces bidirectional connections and weighted feature fusion, which can better utilize the features of multi-scale targets at different levels. Experimental results show that on the DOTA2.0 dataset, the improved algorithm with CBAM and BiFPN improves the mean average accuracy (mAP) by 1.1% compared with Yolov8n; after the sliding window is introduced, the mAP is increased by another 0.3%.