An optimized YOLOv11-based model for surface defect detection on valve stems in refrigeration equipment
TL;DR
The RDD-YOLO model based on the YOLOv11n architecture is presented, which replaces specific convolutions in the backbone and neck with receptive field channel attention convolutions (RFCAConv) to enhance local target perception, and constructs a depth-adaptive kernel spatial pyramid pooling factorization (DAK_SPPF) module featuring DAKConv to modulate multi-scale feature fusion.
Abstract
Surface defect detection on industrial components remains challenging due to difficult feature extraction, low detection accuracy in complex backgrounds, and high computational demands. To address these challenges, this study presents the RDD-YOLO model based on the YOLOv11n architecture. The proposed model replaces specific convolutions in the backbone and neck with receptive field channel attention convolutions (RFCAConv) to enhance local target perception, and constructs a depth-adaptive kernel spatial pyramid pooling factorization (DAK_SPPF) module featuring DAKConv to modulate multi-scale feature fusion. To reduce computational complexity, a lightweight deep dynamic efficient (LDDE) detection head is structured, combining dynamic efficient local convolution (DELConv) and dynamic efficient depthwise convolution (DEWConv) to realize a lightweight Detect_DyHead architecture. The Wise Intersection over Union Version 3 (WIoU v3) loss function is introduced to optimize weight allocation for low-quality samples. Experimental results demonstrated that the proposed model achieved mAP@0.5 improvements of 5.7%, 4.4%, and 2.8% on the public NEU-DET, public DeepPCB, and custom valve stem datasets, respectively, thereby validating its efficacy for industrial defect detection.