The enhanced EME-YOLOv11 for real-time polarizer defect detection

Manual inspection of polarizer defects is time-intensive, and machine vision methods that rely on handcrafted features often reduce accuracy. Deep learning enables automatic feature learning, providing a more efficient solution. This study proposes EME-YOLOv11, an improved model based on YOLOv11, designed to satisfy the real-time requirements of industrial polarizer defect detection. A novel edge information and spatial information enhancement module replaces the cross-stage partial convolution module with two convolutional layers (C3K2) of the YOLOv11 backbone, combining a Sobel-based convolution (SobelConv) branch for edge extraction with a convolutional branch for spatial information, thereby enriching feature representation. In addition, a multiscale shared dilated convolution module substitutes the SPPF structure, enhancing fine-grained feature extraction while reducing trainable parameters. For the network detection head, grouped convolution is employed to construct EfficientHead, which improves scale adaptability, resource utilization, and inference speed. Experimental results demonstrate that EME-YOLOv11 outperforms YOLOv11 in terms of precision, recall, F1 score, FLOPs, and model size, confirming its effectiveness and industrial applicability for real-time polarizer defect detection.