The lack of measured deviation of the strip during cold-rolled process may cause problems of reduction in flatness the control accuracy and strip breakage. An intelligent detection method for coldrolled strip deviation based on machine vision is proposed in this article. A lightweight network structure is constructed, which is based on classic network UNet for intelligent segmentation of cold-rolled strips. The MobileNetV2 is used to replace the original contraction path of UNet and the channel attention ECA_Module is embedded in the connection structure, which effectively reduces the amount of network parameters while enhancing the perception ability of target features. The strip region segmentation model of the cold-rolled strip (SRS_M) can be obtained by training this network. The accuracy indicators mIoU and mPA of SRS_M could reach 98.83% and 99.36%, respectively. The running time of a single image is 40.57 ms. An intelligent detection model of strip deviation is formulated by combining SRS_M and edge position extraction algorithm. The 1 503 deviation sample data are collected through an edge detection device installed on site, which is used to evaluate the proposed method. The results show that the absolute error of 92.82% of the samples is within the range of ±2 mm, and the absolute error of all samples is within the range of ±3.5 mm, which shows the effectiveness of the proposed method in this article.