An Effective Fusion Network for BGA Bubbles Segmentation in X-Ray Image

Abstract

Automatic segmentation of BGA bubbles is of great significance for industrial defect detection. In recent years, convolutional neural network (CNN) has achieved good performance in the field of image processing and is widely used in defect segmentation tasks. Nevertheless, limited by the local receptive field and inductive bias of CNN, existing methods cannot capture sufficient global dependencies for accurate BGA bubble segmentation. Recently, the methods of combining Transformer and CNN can utilize the advantages of both to achieve full mining of sufficient global information and local information. However, the gap between Transformer representations and CNN representations is often ignored in most existing methods, which leads to the problem of feature conflict. In this paper, we propose a novel Effective Fusion Network to merge the gap between Transformer and CNN and to segment small bubbles and bubbles with blurred boundaries. Specifically, we propose a feature alignment module (FAM) to effectively fuse Transformer features and CNN features to merge the gap between them. Furthermore, a consistency constraint loss function is added to the overall loss function for better information complementarity. The experimental results show that our proposed method can more accurately segment small bubbles and bubbles with blurred boundaries on our own collected BGA bubble dataset. Besides, we have achieved the state-of-the-art results with a Dice of 88.92%, an IoU of 80.52% and a Sensitivity of 90.40%