In software testing， determining the image similarity of GUI heterogeneous scenes is the key to improving testing efficiency and quality. The article proposes an image similarity analysis method based on deep neural networks. Firstly， the GUI image was preprocessed and image features were extracted using VGG 16， while an OCR module was introduced to obtain text information. In the similarity calculation stage， the similarity between image features and text was calculated separately. The similarity scores were obtained through cross entropy loss function and string matching algorithm， and finally weighted sum was performed. After verification， this method can effectively resist noise and deformation interference caused by device differences， improve the efficiency and quality of GUI testing， and provide accurate and comprehensive results and decision making basis for testers.