AYOLOv8: Improved Detector Based on YOLOv8 to Focus More on Small and Medium Objects

Object detection has gradually become one of the essential technologies in the field of artificial intelligence. YOLOv8 is a popular version of YOLO series and has attracted the attention of scholars due to its excellent detection performance. However, this model shows slightly inferior detection performance for small and medium-sized objects. To cope with this issue, in this paper, we propose three effective modules and embed them to YOLOv8 for improving its overall detection performance, resulting in a new detector named AYOLOv8 (Augment YOLOv8). Specifically, a fully-dimensional dynamic convolution is firstly designed to enhance the detector’s spatial information extraction ability. Secondly, a new non-local neural networks is introduced to enhance the receptive field of deep features while capturing stronger context dependencies. Lastly, a more reasonable weight allocation for feature fusion is presented to reduce feature redundancy caused by the feature pyramid. Extensive experiments on MS COCO-2017 dataset show that AYOLOv8 significantly improves the performance for detecting small and medium-sized objects and achieves lower model complexity and better detection accuracy than other outstanding methods.