UAV target detection network modeling based on improved YOLOv8

The YOLO-V8 network, extensively utilized in UAV target detection, faces challenges in classification accuracy, localization precision, and detection velocity when confronted with dataset sample imbalance and cross-scale detection issues. In the VisDrone dataset, recognizing tiny objects is a significant challenge, primarily attributed to the model’s limited cross-scale detection capabilities, biased predictability in the presence of data sample imbalance, and excessive model parameter redundancy. To address these issues, this paper introduces the DAT-YOLO-V8, a target detection network model that incorporates five enhancements to YOLOv8: the introduction of SPD-Conv, a deformable convolutional layer, to bolster the extraction of features from small targets; the incorporation of a Deformable Attention mechanism to concentrate on key regions and enhance the precision of small target recognition; the employment of the MDPIoU loss function to refine bounding box learning and enhance localization accuracy; the application of LAMP score-based channel pruning to curtail the parameter count; and the refinement of the pruned model to enhance performance through knowledge distillation within a TeacherStudent model framework with a consistency protocol. Experimental results demonstrate that the DAT-YOLO-V8 network achieves superior detection performance, with a $12 \%$ and $14 \%$ increase in mAP50 and mAP50-95, respectively. Furthermore, it outperforms other state-of-the-art target detection algorithms in accuracy and speed, validating the efficacy and advancement of our enhanced algorithm.