Existing surface defect detection methods for micro-motor commutators suffer from low detection accuracy, poor real-time performance, and high false detection and missed detection rates for small targets. To address these issues, this paper proposes a high-performance and robust commutator surface defect detection model (CLS-YOLO), using YOLOv11-n as the baseline model. First, a lightweight Cross-Scale Feature Fusion Module (CCFM) is introduced to integrate features from different scales, enhancing the model’s adaptability to scale variations and ability to detect small objects. This approach reduces model parameters and improves detection speed without compromising detection accuracy. Second, a Large Separable Kernel Attention (LSKA) module is incorporated into the detection head to strengthen feature understanding and capture, reducing interference from complex surface patterns on the commutator and significantly improving adaptability to various target types. Finally, to address issues related to the center point location, aspect ratio, angle, and sample imbalance in bounding boxes, SIoU Loss replaces the CIoU Loss in the original network, overcoming limitations of the original loss function and enhancing overall detection performance. Model performance was evaluated and compared on a commutator surface defect detection dataset, with additional experiments designed to verify the model’s effectiveness and feasibility. Experimental results show that, compared to YOLOv11-n, the CLS-YOLO model achieves a 2.08% improvement in mAP@0.5. This demonstrates that CLS-YOLO can accurately detect large defect targets while maintaining accuracy for tiny defects. Additionally, CLS-YOLO outperforms most YOLO-series models, including YOLOv8-n and YOLOv10-n. The model’s parameter count is only 1.860 million, lower than YOLOv11-n, with a detection speed increase of 8.34%, making it suitable for deployment on resource-limited terminal devices in complex industrial scenarios.

In response to the current issues of poor real-time performance, high computational costs, and excessive memory usage of object detection algorithms based on deep convolutional neural networks in embedded devices, a method for improving deep convolutional neural networks based on model compression and knowledge distillation is proposed. Firstly, data augmentation is employed in the preprocessing stage to increase the diversity of training samples, thereby improving the model’s robustness and generalization capability. The K-means++ clustering algorithm generates candidate bounding boxes, adapting to defects of different sizes and selecting finer features earlier. Secondly, the cross stage partial (CSP) Darknet53 network and spatial pyramid pooling (SPP) module extract features from the input raw images, enhancing the accuracy of defect location detection and recognition in YOLO. Finally, the concept of model compression is integrated, utilizing scaling factors in the batch normalization (BN) layer, and introducing sparse factors to perform sparse training on the network. Channel pruning and layer pruning are applied to the sparse model, and post-processing methods using knowledge distillation are used to effectively reduce the model size and forward inference time while maintaining model accuracy. The improved model size decreases from 244 M to 4.19 M, the detection speed increases from 32.8 f/s to 68 f/s, and mAP reaches 97.41. Experimental results demonstrate that this method is conducive to deploying network models on embedded devices with limited GPU computing and storage resources. It can be applied in distributed service architectures for edge computing, providing new technological references for deploying deep learning models in the industrial sector.