Abstract
This paper introduces a novel method for removing highlights from the surface of metal ingots, which is a critical pre-processing task for ingot detection and quality assessment. The highlight removal problem is further complicated by the huge area, high intensity, and color similarity to the background, as well as the difficulties of gathering a sufficient number of labeled datasets for network training. To overcome these, this paper proposes a Cycle-GAN network based on double-mask guidance to remove highlight signals from unlabeled metal ingots. This network utilizes double-mask guidance to extract features from both highlight and highlight-free areas in the image. Additionally, a residual attention module and a channel attention module are incorporated to enhance the representation of regional features. To enhance the restoration of texture structural information within the highlight regions, we propose a Texture Recovery Attention Module. This module utilizes extracted features from highlight-free regions to compute similarity attention, effectively combining texture features within the highlight regions and transferring the texture information. For training the proposed network, we introduce a highlight-robust feature perception loss function that supervises the training of the network. Experimental results demonstrate that the proposed method achieves the best result in the task of removing highlights from metal surfaces, preserving more detailed texture information. To verify the generality of the method, we conducted experiments on the SHIQ dataset and achieved better results in the recovery of details. Finally, we verify the effectiveness of each module through ablation experiments.