In the field of online automated defect inspection for small-size liquid crystal display modules (LCMs), the accuracy of module loading is crucial for the subsequent lighting inspection. However, due to the physical characteristics of the module’s flexible ribbon cable, the ribbon often exhibits varying degrees of curling, causing conventional monocular vision systems to frequently encounter local underexposure or overexposure when positioning the workpiece, resulting in loss of local details and significantly affecting subsequent positioning and loading. To address the problem of local image degradation caused by abnormal exposure, this study proposes a regional image generation method based on a dual-layer generative adversarial network (GAN) with multi-scale supervision. This method first uses a mask localization module to restrict the region for image generation, then employs multi-scale local generation and adversarial learning to produce high-fidelity images in areas with local exposure anomalies, and finally uses a newly added global discriminator to regulate the edges of the generated images, allowing the generated images to smoothly connect with the original images, thereby achieving local image repair. Compared to single-layer GAN models, the repaired overall image achieved a peak signal-to-noise ratio (PSNR) improvement of 3.01% and a structural similarity (SSIM) improvement of 11.7%, while the PSNR of the repaired images in exposure-anomalous regions increased by 49.04% and SSIM increased by 23.18%. In addition, not only does the model produce restored images with excellent visual effects, but through verification by deployment on an actual factory production line, the error between the restored images and normal samples is within 0.08 mm, validating the model’s value in practical industrial applications.

In the age of smart technology, the widespread use of small LCD (Liquid Crystal Display) necessitates pre-market defect detection to ensure quality and reduce the incidence of defective products. Manual inspection is both time-consuming and labor-intensive. Existing methods struggle with accurately detecting small targets, such as point defects, and handling defects with significant scale variations, such as line defects, especially in complex background conditions. To address these challenges, this paper presents the YOLO-DEI (Deep Enhancement Information) model, which integrates DCNv2 (Deformable convolution) into the backbone network to enhance feature extraction under geometric transformations. The model also includes the CEG (Contextual Enhancement Group) module to optimize feature aggregation during extraction, improving performance without increasing computational load. Furthermore, our proposed IGF (Information Guide Fusion) module refines feature fusion in the neck network, preserving both spatial and channel information. Experimental results indicate that the YOLO-DEI model increases precision by 2.9%, recall by 13.3%, and mean Average Precision (mAP50) by 12.9%, all while maintaining comparable parameter counts and computational costs. These significant improvements in defect detection performance highlight the model’s potential for practical applications in ensuring the quality of LCD.