Research on Increasing the Robustness of Image Recognition Models Against Different Sets

Research on Increasing the Robustness of Image Recognition Models Against Different Sets

In view of the problem that the recognition performance of the image recognition model trained by a specific dataset is significantly reduced after being transplanted to different dataset, an improved algorithm for attribute combination optimization based on the minimal weighted random search (MinW-Rsearch) and “Equal-Sum” (E-S) judgment method is proposed. First, the MinW-Rsearch algorithm is used to search the image attribute combination and the searched attribute combination is filter through the E-S judgment method. Then, the image that being transformed by the selected attribute combination is input to the improved neural network. Finally, the Adam optimization algorithm is used to train the model. The improved model was transplanted after training, and many experiments were carried out to compare the average classification accuracy of the transplanted model with that of the original model. The experimental results show that the recognition accuracy is increased by at least 5% after the improved digital recognition model is transplanted, which obviously improves the robustness of the recognition performance after the model is transplanted.