GRIN Lens End Face Classification Detection Based on Deep Learning

Abstract

With the rapid development of optical communication technology, the demand for gradient-index (GRIN) lenses has increased significantly, making quality inspection of lens end faces an increasingly critical issue. In particular, accurate detection of defects on both ends of GRIN lenses remains a challenging task. To address this problem, this study employs a transfer learning-based parameter fine-tuning approach to evaluate the classification performance of four deep learning models on a defect dataset. Among the evaluated models, ResNet50 and DenseNet-169 demonstrated superior performance and were selected for further optimization. Attention mechanisms, including squeeze-and-excitation (SE) and convolutional block attention module (CBAM), were incorporated into these models to enhance feature representation. Experimental results show that, after integrating the SE module, the classification accuracy of ResNet50 and DenseNet-169 increased by 0.0243 and 0.0272, respectively. With the addition of the CBAM module, the accuracy improvements reached 0.0437 for ResNet50 and 0.0506 for DenseNet-169. These results indicate that the proposed improvements significantly enhance the defect detection capability of the models. All evaluation metrics show consistent improvement over the baseline models, demonstrating that the integration of attention mechanisms effectively increases the classification accuracy and overall performance of the original network architectures.