YOLOv8 Improved Model for Small Target Recognition

Small target recognition and aerial photography detection have a vast development prospect in civil and industrial fields. Nowadays, deep learning is developing rapidly, which provides more effective solutions for small target recognition, but factors such as lighting conditions, mutual occlusion of small objects in the image, and changes in shape and size bring challenges to detection accuracy and real-time detection. This paper proposes a DSL-YOLOv8 detection model, which owns excellent robustness and the ability to balance detection accuracy and speed. First, utilizes the DCN (Deformable Convolution Networks) structure to improve all C2f (CSP Bottleneck with 2 Convolutions) module in the YOLOv8 baseline model, which suppresses the information loss of network depth features during transmission; then, the SPD-Conv (Space-to-Depth) module is used to improve the extraction of fine-grained information by the model; secondly, the LSKA (Large Strip Kernel Attention) mechanism is introduced at the neck part to solve the occlusion of small strip-shaped targets and the leakage detection of sparse small targets; finally, the CIoU (Complete-IoU) loss is replaced by WIoUv3 (Wise-IoU) loss, whose dynamic focusing mechanism enables the detector to dynamically adjust the loss weight according to the quality of the anchor box, thus improving the accurate localization of the bounding box. The performance of the algorithm is evaluated on the widely used VisDrone2019, and ablation experiments, comparative experiments, and generalization experiments were designed to verify the high performance of the DSL-YOLOv8 model. The results show that the DSL-YOLOv8 model has achieved obvious advantages and provides a new solution for small target recognition and detection.