A Hybrid Transformer–CNN Network Based on DINOv3 for Lens Surface Defect Detection

Abstract

Optical lens surface defects are often characterized by low contrast, weak textures, blurred boundaries, and dense distributions, making accurate detection challenging under complex industrial conditions. To address these issues, this paper proposes DWTF-Net, a DINOv3-based Transformer-CNN network for optical lens surface defect detection. The proposed framework contains two core components. First, a Multi-Scale Wavelet Feature Enhancement Block (MSWFEBlock) enhances weak defect features by jointly exploiting spatial-domain structures and frequency-domain details. Second, an MGCBlock-based convolutional branch captures multi-scale local features to complement the global semantic representations extracted by the self-supervised pretrained DINOv3 Transformer. Experiments were conducted on the self-constructed Optical Lens Surface Defect Dataset (OLSD). Under the same training and evaluation settings as the comparison models, DWTF-Net achieved an mAP@50 of 72.6% on OLSD, outperforming the strongest baseline by 5.7 percentage points.