WDS-YOLO: a small object detection algorithm based on YOLOv8s

Small object detection faces three prevalent challenges: feature degradation of tiny objects, ambiguous localization due to dense distribution, and interference from complex backgrounds. The paper introduces WDS-YOLO, a computationally efficient detection framework modified from YOLOv8s. A key innovation involves implementing a reduced-size prediction head to replace the default bulky configuration, resulting in enhanced small-object recognition capabilities. Second, to mitigate the degradation of fine-grained details in deep networks, wavelet transform convolution is incorporated to separate and enhance high-frequency texture features within the frequency domain. Additionally, the SPDConv (Space-to-Depth Convolution) spatial-channel rearrangement down-sampling module is employed to alleviate information loss. Furthermore, a Spatial Context-Aware Module (SCAM) is proposed, which robust suppression capability against cluttered background artifacts through a dual attention mechanism encompassing both channel and spatial dimensions. Validated on the VisDrone2019 dataset confirm that WDS-YOLO, with only 10.7M parameters, achieves 46.8% mAP50 and 29.3% mAP50:95. This represents an improvement of 8.8% and 5.8% over the baseline YOLOv8s, respectively. Concurrently, the computational cost is optimized to 40.1 GFLOPs. This method offers an efficient solution to the small object challenges