Image processing and deep learning based analysis of 3D X-ray PCB images

Abstract

In this project, I propose a modified U-Net architecture for segmenting PCB (Printed Circuit Board) images. The proposed model consists of an encoder and a decoder structure with a connection of skip that enable integrations of low-level and high-level features for accurate segmentation. To enhance the segmentation performance, I introduce dilated convolutions, dense connections and convolutional layers in the decoder part. Additionally, we adopt a mixture of binary cross-entropy as well as dice loss functions to optimize the model during training. The intended model is assessed on the public dataset of PCB images. Comparative analysis reveals that our model’s performance surpasses that of its competitors with an general segmentation accuracy of 94.2%. Furthermore, the proposed model is computationally efficient and can segment a PCB image in 1.52s.