Identifying the Working Conditions of Semisolid Metal Slagging Operations Under Strong Light Based on Gabor Feature and Multichannel PGAN

In this paper, a detection framework is proposed that combines strong light removal, working condition identification, and slagging quality evaluation. This framework is applied to a semisolid metal solution slagging operation system, which is an approach that was not available in previous work. Due to the strong time-varying nature and texture of metal solutions, it is difficult to identify different oxidized slag working conditions. The Gabor transform is used to model images of semi-solid solutions. The t location-scale distribution and extreme value distribution are used to fit the statistical characteristics of Gabor responses. A least squares support vector machine (LS-SVM) classifier is trained to recognize the model based on the statistical feature vectors. This enables working condition identification and quality evaluation. To solve the strong light interference on the semisolid metal solution, color space conversion and a parallel generative adversarial network (PGAN) is proposed to restore the missing texture details in the light area and improve the image quality of metal solutions. In the experiments, the accuracy of working condition identification and quality evaluation reached 95%, and the product quality met the production requirements. Note to Practitioners —This article is motivated by the problem of ensuring the quality of slag removal by robots in order to completely remove oxidized slag from the surface of each metal ingot.To guarantee the quality of each metal ingot, the production line uses two robots to remove oxide slag. This is accomplished in two stages: removal of coarse and fine slag. Before coarse slag removal, a visual inspection framework is used to recognize work conditions and calculate the shovel depth for the coarse slag robot. After the coarse slag removal, the ingot surface is assessed for quality. If it is not qualified, fine slag removal is performed to remove the residual oxidized slag. Additionally, a light removal algorithm based on an improved generative adversarial network is proposed to improve the quality of the metal solution image. This detection framework has been applied in actual production, and future research will focus on the visual servo control of the slag-picking robot.