Research On Robustness Of Deep Convolutional Neural Networks For Computer Vision Tasks

In recent years,with the rapid development of deep learning technology,deep learning has become a research hotspot and mainstream of artificial intelligence.Although some vision applications based on deep learning technology have begun to develop into products,there are still many limitations in the comprehensive and in-depth transformation of deep learning technology into the industry.Among them,the robustness of deep models is one of the key factors.The deep model robustness requires the model has significant performance on regular training sets,good stability on noisy datasets,and good generalization performance on test data with shifted distributions.In order to improve the robustness of the model in different vision test scenarios comprehensively,this dissertation firstly analyzes the model robustness problem and uses the modeling result as a unified basis for evaluating the robustness of the model.Then this dissertation takes the computer vision task as the main application scenario,combined with the limiting factors of the robustness of the deep model,and conducts a modular analysis of the deep model.According to the analysis results,this dissertation takes enhancing the feature extraction and identification capability of the model as the core entry point and conducts in-depth research from three aspects: enhancing the feature extraction capability of the convolutional neural network,the feature identification capability of the model and the consistent response of the features in the model.Finally,this dissertation has achieved three innovative results,which are:(1)An adaptive dilated convolutional neural network based on online reasoning strategy is proposed,which realizes that the model adaptively adjusts the receptive field of the convolutional kernel during the training process,and enhances the feature extraction ability of the convolutional neural network.Dilated convolution makes the convolutional kernel obtain a larger receptive field by introducing the parameter of dilated value,while the wrong selection of dilated value will reduce the efficiency of dilated convolution and affect the model training.Aiming at this problem,this dissertation proposes a dilated value online reasoning strategy,and approximates the dilated value sampling process by introducing the Gumbel-Softmax function.At the same time,various inter-layer information aggregation modes such as Markov aggregation mode are designed to model the information update process of hidden units.The experimental results show that the adaptive dilated convolutional neural network proposed in this dissertation can be flexibly embedded in multiple deep models,and bring a stable improvement to the robustness of the model on typical vision tasks.(2)The Active Dropblock method based on active learning is proposed,which realizes the online optimization of the mask in Dropblock and enhances the model’s ability to identify the effective features contained in the data.Dropblock overcomes the low efficiency problem of traditional Dropout in convolutional neural networks by using structured masks to perturb the features.However,the generation of masks lacks reasonable guidance,which imposes burden on model training efficiency.To solve this problem,this dissertation firstly optimizes the update process of the model based on active learning,and then realizes the online optimization of the mask by constructing a parameter selector.The experimental results show that the Active Dropblock proposed in this dissertation can effectively improve the model’s resistance to adversarial examples,and at the same time,this method improves the model’s data utilization efficiency and the activation response of features in the model.(3)A novel scheme towards enhancing deep model accuracy and robustness is proposed,which improves the generalization ability of the model on unknown testing data and enhances the consistent response of features in the model.The data augmentation technology can expand the data distribution of the training set by adding the enhanced image data to the training set,and improves the generalization ability of the model on unknown test data to a certain extent.However,limited data augmentation operations cannot cope with complex data in realistic testing environment.To address this problem,this dissertation first constructs an enhanced variant of clean images by introducing different data augmentation methods,and uses these image samples to construct classsensitive positive and negative sample pairs,and then input these samples into the model for training simultaneously.Then this dissertation proposes a feature consistency constraint strategy and a prediction distribution consistency constraint strategy based on contrastive learning,which constrains the consistency expression of features and prediction distributions.The experimental results show that the proposed scheme can effectively enhance the consistent response of features in the model and the efficiency of data utilization.