Obtaining a large number of unqualified product samples in industrial production is an arduous task. It is challenging to learn the features of few-shot object images. Despite the limited number of original images, we developed a transfer learning method called LDFISB (Large-scale Dataset to Few-Shot Image with Similar Background) that provides a feasible solution. LDFISB is trained on a large-scale dataset such as CIFAR100, and then the model is fine-tuned based on the original model and parameters to achieve classification tasks on a new APSD (auto part surface dataset). Batch normalization, padding, and Weighted Cross Entropy Loss are employed in the training processes. Hyper-parameters are configured according to Hyper-table to enhance the accuracy of the prediction. The CIFAR10, CIFAR100, and ImageNet were considered as pre-training datasets, and the LDFISB method is capable of accurately predicting the flaw area of the product image. The LDFISB method achieves the highest accuracy on the CIFAR100 pre-training dataset.