The Enhanced Network Swin-T by CNN on Flow Pattern Recognition for Two-phase Image Dataset with Low Similarity

Abstract

In the two-phase flow experiments with different conditions of materials and process parameters, the collected image dataset with the low similarity and small amount was difficult for the common deep learning algorithms to achieve a high-precision recognition of flow pattern. due to the low extraction capability of global features. In this article, we proposed a new deep learning algorithm to enhance Swin-T network by CNN which combined the advantages of Swin-T network with Dynamic Region-Aware Convolution. The new algorithm retained the window multi-head self-attention mechanism and added the self-attention adjustment module to enhance the extraction of image features and the convergence speed of network. It significantly improved the recognition accuracy of the different flow patterns in the sharp and blurred images. The enhanced network Swin-T by CNN had the high applicability to the classification of image dataset with low similarity and small amount.