YOLO-DC: Enhancing object detection with deformable convolutions and contextual mechanism

Existing object detection methods often optimize model structure, loss functions, and data preprocessing, but enhancements via convolutional techniques are often overlooked. Additionally, increasing network depth leads to the loss of key features. This paper introduces YOLO-DC, which enhances object detection with deformable convolution and contextual mechanisms. YOLO-DC includes a Deformable Convolutional Module (DCM) and a Contextual Fusion Downsampling Module (CFD). The DCM module uses a modified deformable convolution of Multiscale Spatial Channel Attention (MSCA) to extend the receptive field and improve feature extraction. The CFD integrates contextual and local features post-downsampling, reducing information loss. Compared to YOLOv8-N, YOLO-DC-N improves average precision (AP) by 3.5%, reaching 40.8% on the Microsoft COCO 2017 dataset, with similar FPS and inference time. The model demonstrates superior performance compared to other state-of-the-art detection algorithms. The source code is available at: https://github.com/Object-Detection-01/YOLO-DC.git.