Small dense Mini/Micro LED high-precision inspection based on instance segmentation with local detail enhancement

Abstract

As Mini/Micro LED size decreases and integration proliferates, Mini/Micro LED industrial image chips are small in size and densely distributed, posing a huge challenge for high-precision mass inspection. Instance segmentation (IS) networks providing pixel-level localization and recognition are potential for fine inspection of mass Mini/Micro LEDs. However, for small and dense targets, generic IS suffers from the lack of localized detail features, a less specificity of deep and shallow features fusion, and the loss of key details in the upsampling. Therefore, an IS framework for local detail enhancement is proposed to enhance local features from three parts: feature extraction backbone, feature fusion, and upsampling. To improve the local detail extraction, a local aggregation backbone based on multiscale receptive fields is proposed by increasing the weights of small receptive fields to extract and retain more local detail features. For adequate feature fusion, a category hybrid mask fusion module is proposed to guide local feature fusion with the global information of categories. Adaptive hybrid upsampling is introduced as an alternative to traditional uniform upsampling. It is designed to restore high-resolution detail information and reduce computation by filtering keypoints, oversampling keypoints and undersampling residual points. For the Mini/Micro LED dataset collected from the production line, mAP50, mAP75, and mAP90 are 97.38 %, 92.07 %, and 81.72 %, respectively, which are higher than the existing object detection and IS methods.