Image recognition of limited and imbalanced samples based on transfer learning methods for defects in welds

Abstract

Welding quality inspection is critical to the quality control of the welding structure. Traditional manual detection requires experienced workers and the method is time-consuming. Currently, deep learning has made great progress in the field of image recognition. However, in terms of industrial defect detection, the contradiction between huge computational parameters and limited imbalanced samples still exists, which makes the deep learning method unable to play its role to the maximum extent. In this case, a combination of deep learning methods and transfer learning methods is considered to improve the performance of the model on limited and imbalanced data sets. The VGG16 model, which is pre-trained on a large number of source data sets, is fine-tuned on our target data sets. In this paper, a new DTN-VGG16 structure is proposed, in which a Global Average Pooling (GAP) layer, Batch Normalization (BN) layer, and Soft-max classifier are added on top of the frozen pre-training model, which can effectively reduce the number of parameters and accelerate the convergence speed of the model. In addition, focal loss function is used to replace the commonly used multi-class cross entropy loss function. By adjusting the parameters of the loss function, the training process has better convergence. Experimental results show that DTN-VGG16 has better performance than other traditional machine learning methods and deep learning models. And the proposed model has good robustness and generalization performance when valid on the real data of aerospace welding seams which is difficult to learn.

A study on spot welding quality judgment of stainless steel plate based on semi-supervised conditional generation adversarial network

“…There are many methods available to deal with unbalanced datasets, such as under-sampling, oversampling, and the generative method. For example, to improve the performance of the model on limited and imbalanced data sets, Zhang et al 1 proposed deep learning and transfer learning methods. Aiming at the class imbalance, Wang et al 2 proposed a new graph-based method, named anchor-based class-balanced GCN (ACB-GCN).…”