LCSeg-Net: A low-contrast images semantic segmentation model with structural and frequency spectrum information

Abstract

Semantic segmentation in low-contrast images is a challenging problem due to the ambiguous boundary of the segmented target and indistinguishable noise. Current models generally segment the target with local features, which leads to the lack of structural information, which reduces the performance. And they directly fuse multi-scale features to keep details, while it may integrate noise into the fused features. These problems reduce the performance of segmentation in the low-contrast image. To solve the above issues, we propose an image semantic segmentation model called Low-Contrast Segmentation Net (LCSeg-Net). The model enhances the structural information with the global context and reduces the noise of the fused features through the adaptive fusion way. Meanwhile, to improve the robustness of LCSeg-Net, we augment the low-frequency spectrum of the fused feature in the training phase. Extensive experiments are conducted on the five public datasets. Results show the proposed model achieves the best comprehensive performance in the low-contrast image.