| Image detection and classification is the most basic task in the field of CV.Automatic object detection and classification in images is a prerequisite for many advanced intelligent tasks based on machine vision.At present,the image detection and classification method based on deep neural network is the most advanced and most concerned technical scheme in this field.However,the deep neural network has serious over-parameterization problems,which means that the image detection and classification methods based on the deep neural network will cost a lot of computing,storage,power,and other resources.In many industrial scenarios,such as autopilot,Internet of Things,and space-borne remote sensing,not only computing,storage,power,and even communication resources are extremely limited,but also multiple intelligent applications often need to be deployed at the same time.Therefore,the contradiction between the huge resource requirements of image detection and classification deep neural networks and the resource constraints of edge computing devices has become a bottleneck that restricts the development and promotion of intelligent applications based on deep learning technology.In order to alleviate the contradiction between the deep neural network and the hardware resources,this dissertation summarizes several scientific problems that need to be solved by analyzing the advantages and disadvantages of the existing deep neural network optimization methods.In addition,to eliminate the over-parameterization problem of the deep neural network,this dissertation conducts in-depth research on image detection and classification depth neural network optimization methods.The research contents and main contributions of this dissertation include the following aspects:1.Aiming at the problem of the low accuracy of the existing heuristic pruning criteria and the blind area of traditional pruning strategies for complex network pruning,a deep neural network pruning method for image classification based on the feature map information quantity criterion is proposed.According to the analysis of the relationship between the importance of filter parameters and the output feature map,the proposed method establishes the filter parameters’ importance evaluation criteria based on the information quantity of the quantized feature map,which improves the classification accuracy of the pruned image classification deep neural network.The proposed method also establishes a filter pruning strategy for convolution layers in the pruning blind area by analyzing the coupling relationship between different convolution layers of the complex structure network,which significantly improves the compressible rate of the complex structure deep neural network.2.Aiming at the generalization problem of parameter importance evaluation criteria,an image classification deep neural network optimization method based on feature image quality criteria is proposed.This method establishes the filter importance evaluation criteria based on the feature map’s quality by analyzing the possible commonality between the effective parameter importance evaluation criteria,which proves the feasibility of generalizing from a parameter importance evaluation criteria to a kind of parameter importance evaluation criteria.Based on the proposed feature map quality criterion,the method also establishes the image classification deep neural network pruning method and knowledge distillation method for network optimization,which proves the feasibility of generalizing from one parameter importance evaluation criterion to another network optimization method.3.To address the problem that the traditional network pruning method is challenging to work because the parameters of deep neural networks for image detection are more sensitive than image classification,an optimization method for image detection deep neural network based on the characteristics of remote sensing images is proposed.After analyzing the characteristics of the image detection deep neural network and the design defects of the traditional network pruning pipeline,the proposed method optimizes the traditional network pruning pipeline by establishing a four-stage network pruning process that only needs to be executed once.According to the characteristics of remote sensing images,the proposed method also establishes a pruning proportion adjustment scheme,which improves the optimization effect of the deep neural network for object detection in remote sensing images.4.Aiming at the lack of objectivity of heuristic pruning criteria,an object detection deep neural network optimization method based on a multi-objective optimization algorithm is proposed.The proposed method completely discards the heuristic network pruning criterion but models the image detection network optimization problem as a multiobjective optimization problem,and uses a multi-objective optimization algorithm to solve the optimized network structure.Finally,the proposed method will output the optimization scheme set that satisfies different balance factors among multiple objective functions.Most importantly,this method realizes objective image detection deep neural network pruning.The above work has been validated on various image classification and detection data sets.The experimental results show that the proposed methods can effectively reduce the cost of hardware resources cost by deep neural network inference while maintaining the performance of image classification and detection deep neural network.This dissertation also provides a technical solution for deploying intelligent technology based on the deep neural network on the resource-constrained edge computing platform. |
