RDD-YOLO: A study on an improved valve stem surface defect detection algorithm for YOLOv11

Addressing challenges in detecting surface defects on refrigeration equipment valve stems—including difficulty in extracting small target features, low accuracy against complex backgrounds, and high computational demands—this paper proposes an improved RDD-YOLO model based on YOLOv11. First, the model replaces part of the convolutions in the backbone and all convolutions in the neck with receptive field channel attention convolutions (RFCAConv), enhancing local perception and channel attention to improve feature extraction for small target defects. Second, it constructs a depth-adaptive kernel spatial pyramid pooling factorization (DAK_SPPF) within the backbone network and designs depth-adaptive kernel convolution (DAKConv) to optimize multi-scale feature fusion, thereby improving detection accuracy in complex backgrounds. To reduce parameters and computational complexity, a lightweight deep dynamic efficient (LDDE) detection head is proposed. Combining dynamic convolution and the efficient local attention (ELA) module, a lightweight Detect_DyHead architecture is designed. Finally, the wise intersection over union version 3 (WIoU v3) loss function is introduced to optimize weight allocation for low-quality samples, enhancing recognition and generalization capabilities for such instances. Experimental results demonstrate that the improved model achieves a 5.7% and 1.6% increase in mAP@0.5 on the NEU-DET and custom datasets, respectively, compared to the original model. Concurrently, computational complexity is reduced by 6.4%, while detection precision and recall improve by 7.9% and 5.0%, respectively. This effectively enhances the detection performance of valve stem surface defects, meeting the high-precision and real-time detection demands of industrial scenarios.