Weakly supervised aluminum alloy microstructure image semantic segmentation based on consistency constraint

Abstract

To improve the quality of alloys, experts need to intelligently analyze the metallographic images, where semantic segmentation serves as a vital means of achieving this analysis. However, traditional fully supervised semantic segmentation methods often suffer from high annotation costs. Weakly supervised semantic segmentation methods require only simple image-level annotations to achieve high-precision semantic segmentation. This paper constructs a dataset of weakly supervised microstructure images of aluminum alloys, greatly reducing manpower and time costs. Furthermore, addressing the sparsity and incompleteness of Class Activation Maps (CAMs) in traditional weakly supervised methods, a weakly supervised semantic segmentation method for aluminum alloy microstructure images is proposed based on consistency constraints. The method first introduces consistency constraints in the feature extraction process to improve the quality of CAMs. It then uses the improved CAMs as a target cue to generate affinity labels and employs the AffinityNet network and random walk strategy to mine target pixels in non-discriminatory areas, obtaining completer and more detailed pixel-level pseudo labels. Finally, semantic segmentation of the obtained pixel-level pseudo labels is performed using the U-Net network to obtain the final segmentation result. Extensive experimental results demonstrate that our method effectively improves the quality of CAMs, thereby enhancing the performance of the segmentation network.