Chip dies defect detection model based on MobileDDD-Net

This paper introduces a novel lightweight model for wafer defect detection designed to overcome the high time and space complexity of traditional deep learning methods, which hinder deployment in embedded environments. The proposed model utilizes deep separable convolution, the mish activation function, and Gaussian pooling to improve detection speed, and accuracy, and reduce storage space. Experimental results demonstrate that this model outperforms MobileNet-V3, MobileNet-V2, and VGG-16 in terms of speed, storage efficiency, and accuracy, achieving a remarkable 99.72%. Even at a lower resolution of 128 pixels, the accuracy only slightly decreases to 99.65%, indicating potential for further model size reduction.