Multi-Scale Feature Fusion Based on Improved YOLOv8 Networks for Visual Relationship Detection

Despite its high efficiency in real-time detection, YOLOv8 faces limitations in dense scenes with many overlapping objects. Its deficiencies in multi-scale feature handling and deep feature extraction affect the recognition of subtle scene differences. This paper proposes RelYOLOv8, an enhanced visual relation detection model based on YOLOv8 network, aiming to improve object detection accuracy and visual relation reasoning in complex scenes. It replaces YOLOv8's original C2f module with the C2f_Lite module, combining standard and depthwise separable convolution to optimize computational efficiency, reduce the load, and enhance detection performance, especially for small objects. The Atrous Spatial Pyramid Pooling (ASPP) module is added to better handle multi-scale features, and the Slide Loss function is integrated to strengthen the model's ability to distinguish difficult samples. Experimental results on the VRD dataset show a marked increase in detection accuracy from 0.716 to 0.861, with a 14.5% improvement in accuracy and a 14.2% reduction in missed detection rates. Moreover, RelYOLOv8 clearly outperforms Faster R-CNN and YOLOv8 in complex scenarios.