YOLO-DKR: Differentiable architecture search based on kernel reusing for object detection

In recent years, neural architecture search (NAS) has been used to search high-performance convolutional neural networks (CNNs) for object detection. However, existing NAS methods need expensive computational costs. To address this issue, this paper proposes YOLO-DKR (differentiable architecture search (DARTS) based on kernel reusing for YOLOv7). To reduce GPU memory and search time during NAS training, we employ DARTS based on kernel reusing technique, which shares the weights of candidate convolutions on a single edge and merges them into a single fused convolution. Additionally, we design supernets based on YOLOv7-tiny and YOLOv7 networks, using a joint search strategy for optimal solutions. Finally, during retraining phase, we introduce the coordinate attention (CA) modules to further enhance the model's ability to detect small objects. Experiments demonstrate that our method achieves exceptional search efficiency and accuracy. Our lightweight model, YOLO-DKR-tiny, requires only 0.4 GPU-days, two-thirds of MAE-DET-L's search time, and outperforms most NAS-based methods in search efficiency. When transferred to MS COCO 2017 dataset, YOLO-DKR-tiny surpasses DeepMAD(FCOS) by 0.3% in [email protected]:0.95 with fewer parameters. For small object detection (AP${}_S^{val}$), YOLO-DKR achieves 36.3%, exceeding YOLOv10-C by 0.5%. These results demonstrate our method's effectiveness in achieving superior detection with optimized computational resources.