Deep Convolutional Neural Network Transfer Learning Optimization Based on Visual Interpretation
Yibo Xu, Jiongming Su, Fengtao Xiang, Ce Guo, Haoran Ren, Huimin Lu · 2021
In image classification tasks, the training of deep convolutional neural networks generally requires a large amount of data, and due to the constraints of environment, resources and time, it is of great practical importance to use fewer training samples to obtain a higher recognition rate in the shortest possible time. A deep convolutional neural network transfer learning optimization method based on visual interpretation is proposed for a specific image classification task. Firstly, we use class activation mapping visualization as a visual interpretation, output the class activation heat map of the validation set images, and analyze the reasons for misrecognition of the images. Secondly, we introduce “feedback” by pre-recognizing and visualizing the optimized dataset with the model trained on the original dataset, selecting the images that have a greater impact on improving the recognition rate, and maximizing the impact of the optimized images on the original model. Finally, the model is retrained on the optimized training set. The experimental results show that this method can effectively improve the recognition rate of the transfer learning model for image classification.