Applicable for die surface defect inspection, dicing street defect inspection, RDL defect inspection before/after wafer dicing in advanced semiconductor packaging, as well as bump height and coplanarity measurement.

Deep learning AI algorithm

Deep learning algorithms enable neural networks to automatically learn features and extract information from massive data through training, performing intelligent analysis on image or video data. The current software integrates five universal algorithms based on CNN and Transformer architectures: classification, segmentation, object detection, character recognition, and unsupervised learning (good product learning). These algorithms demonstrate significant effectiveness in practical applications such as defect recognition in complex environments, component classification, and character reading scenarios. With deeper research, we have not only refined existing model architectures to reduce parameter quantities but also leveraged NVIDIA's efficient deep learning inference engines to implement computational graph optimization, accuracy improvement, layer simplification, multithreading, and batch inference processing. These enhancements have substantially improved model inference speed and efficiency.