A Robust Positioning Method Based on Semantic Segmentation Network for DIE Chips

The accurate and fast positioning of DIE chips is crucial for conducting high and low temperature online performance tests during the manufacturing process of chips. The original online performance test equipment for DIE chips cannot accurately locate polluted chips, resulting in serious missing and false detections. To address this issue, this paper proposes a robust positioning method based on semantic segmentation network for DIE chips. Unet is used to segment the two circular electrodes of the DIE chip. After that, image processing techniques, such as contour extraction and circle fitting, are employed to extract the precise location of the DIE chip. To satisfy the requirements of fast and precise positioning, improved strategies such as modifying convolutional blocks, compressing the channels of convolutional layers, the weighted combination of the dice and the crossentropy loss function, and so on are adopted to improve Unet (named IUnet). Additionally, to address the issue of insufficient samples for the low temperature performance test, model-based transfer learning is employed to improve the segmentation accuracy of IUnet. The experimental results show that the proposed method, compared with mainstream methods such as traditional methods (template matching and Hough circle detection), target detection methods (SSD, YoloV7, and FasterRCNN), and semantic segmentation methods (SegNet, DeeplabV3+, Unet, Res-Unet++, and Swin-Unet), has better positioning performance and robustness. Under the premise of achieving a positioning beat of 0.5 s per chip, the positioning accuracy rate of DIE chips can reach 100 %, thereby satisfying the positioning requirements of the enterprise.