The positioning accuracy of the PCB during processing depends on the quality of the MARK point images. The collection of MARK point images is affected by factors such as background, illumination etc., so the classification of images is the key to improve the accuracy of PCB positioning. In this paper, a multidisciplinary design modelling method for product quality is proposed. A classification model through transfer learning based on the ResNet50 network and weights is built. It is verified by a set of customized experiment data that the accuracy of MARK point image classification has reached 98.53%. Compared with traditional classification methods, the accuracy rate of this method is 20% higher, and is more suitable for custom small data sets. It provides a guarantee for the subsequent classification and segmentation of MARK point images with different characteristics.