Optimization of Few-Shot IC Chip Defect Detection Model Based on YOLOV11

With the rapid development of the semiconductor industry and machine vision inspection technology, defect detection technology based on deep learning methods has gradually become the main method used in industrial fields due to its efficiency and accuracy. In order to solve the problems of low detection efficiency and accuracy due to the insufficient sample size of defect datasets in real industrial production, a defect detection model optimized based on the YOLOV11 algorithm framework, called YOLOV11-H, is proposed. The hallucination network implemented based on a multi-layer perceptron (MLP) can learn class-shared features from base classes with rich samples and generate additional hallucination samples for new classes to train the model. The experimental results demonstrate that the improved defect detection model achieves good performance in few-shot scenarios. In the field of IC chip defect detection, the proposed method achieves performance comparable to existing object detection approaches utilizing 100 training samples, while requiring only 2 hallucinated generated samples. When augmented with 5 novel hallucinated samples, the method attains an average precision (AP) of 0.674 across intersection-over-union (IoU) thresholds ranging from 0.5 to 0.95 between predicted and ground-truth bounding boxes. This approach provides an effective solution for few-shot IC chip defect detection tasks.