Improved model MASW YOLO for small target detection in UAV images based on YOLOv8

The present paper proposes an algorithmic model, MASW-YOLO, that improves YOLOv8n. This model aims to address the problems of small targets, missed detection, and misdetection of UAV viewpoint feature detection targets. The backbone network structure is enhanced by incorporating a multi-scale convolutional MSCA attention mechanism, which introduces a deep convolution process to aggregate local information. This method aims to increase small-target detection accuracy. Concurrently, the neck network structure is reconstructed, with the fusion effect of multi-scale weakening of non-adjacent levels addressed by using an AFPN progressive pyramid network to replace the PANFPN structure of the base model. The MSCA and AFPN form a multiscale feature synergy mechanism, whereby the response values of MSCA become inputs to AFPN, and the multiscale integration of AFPN further amplifies the advantages of MSCA. The use of flexible non-maximum suppression Soft-NMS is chosen to replace the non-maximum suppression NMS to improve the model's detection of occlusion and dense targets. The loss function of the model is optimised through the implementation of Wise-IoU, which serves as a replacement for the loss function of the baseline model, thereby enhancing the accuracy of bounding box regression, especially perform better when the target deformation or scale change is large. Experiments conducted on the VisDrone2019 dataset demonstrate that the average detection accuracy of the MASW-YOLO algorithm is 38.3%, which is augmented by 7.9% through the utilisation of the original YOLOv8n network. Furthermore, the number of network parameters is reduced by 19.6%.