DYOLO: A Novel Object Detection Model for Multi-scene and Multi-object Based on an Improved D-Net Split Task Model is Proposed

This paper proposes a novel network model named DYOLO, aimed at improving the accuracy and real-time performance of object detection tasks. This model combines the dynamic feature fusion capability of D-Net with the high-speed detection performance of YOLOv8n. We designed a new convolutional module, DConv, by incorporating the dynamic large convolution kernel (DLK) and dynamic feature fusion (DFF) modules from the D-Net network structure, and applied this module to improve the C2f modules in the backbone and Neck of YOLOv8n. These improvements enable the model to more effectively capture and fuse multi-scale features, enhancing the utilization of global contextual information. Experimental results demonstrate that the improved model achieves higher accuracy and better adaptability in object detection tasks across different scenarios, as verified on public datasets such as NWPU VHR-10 and RSOD, as well as our self-built Car dataset. Furthermore, while maintaining the original real-time performance of YOLOv8n, the model enhances the capability to capture both local details and overall scene information. This study provides a novel solution for object detection tasks in various scenarios.