Research On Deep Learning Model Understanding Method For Object Recognition

Deep learning model has been widely used in various fields of social production,and has achieved remarkable performance in many visual tasks,such as target recognition and classification.However,the end-to-end learning mode makes the model behave as a black box.The opacity and complexity of the model are far beyond the scope of human understanding.Besides the final network output,it is difficult to understand the prediction logic hidden in the network.When the generalization performance of the model is very good,it is difficult to explain why the model makes this prediction and what kind of judgment it is based on.However,the performance of the model in practical application is not perfect,and there will be some mistakes.The cognitive level of the model is relatively low,which limits its further development and application.Many scenarios need to explain the decision-making of the model,especially in some areas with high security requirements.Therefore,the understanding of deep learning model is of great significance and practical value.By analyzing the training process of the convolution neural network model,it is found that the convolution operation can retain the spatial and semantic information of the features in the input samples,and these features can effectively reflect the area of the input samples that the model focuses on.From this point of view,the method of class activation mapping of acc-cam is proposed.The class activation mapping map obtained by this method can make a decision on the category area in the input sample,so that the model can be differentiated according to the different characteristic areas of the input sample in the classification task,so as to more intuitively and conveniently explain the impact of different areas in the sample data on the model classification results Ring.In the aspect of interpretation of convolutional neural network classifier,the contribution degree of features in multi data sample space to classification result judgment is analyzed,and it is found that different features have different contribution degree to classification result judgment.From this point of view,this paper proposes an interpretation tree method based on contribution degree.By combing the feature set extracted from the whole sample space by convolution neural network model,the contribution distribution matrix of feature variables in the data sample space is constructed,and the interpretation tree is generated from the contribution distribution matrix.The interpretation tree can reflect the importance of feature combination of the model in the process of category analysis of specific samples.Using the method of interpretation tree can help people to effectively understand the prediction mode of deep learning model,and make a reasonable explanation for the specific sample data,so as to understand the deep learning model more effectively.In a word,the method of class activation mapping and the method of explanation tree based on contribution degree can help people understand the deep learning model effectively and use it better.