Research On Neural Architecture Search And Application Based On Pixel Difference Convolution

Neural Architecture Search(NAS)is a technology for automatically designing artificial neural networks.Its goal is to learn an artificial neural network topology to achieve optimal performance on target tasks.Recently,neural network structure search technology has attracted more and more attention and research focus in academia due to its potential applications in deep neural network compression,convolutional neural network interpretability,deep learning popularization,and AI antitrust.This paper firstly investigates the historical context and main technical characteristics of neural network structure search technology,and summarizes the current NAS development into search space,search strategy,performance evaluation strategy,etc.The performance of tasks on image classification is visually displayed in the form of organized charts.Based on the research results and thinking,it summarizes the bottleneck problems of the current NAS technology development and the future development direction,in order to provide some help to interested researchers.The core technology of neural architecture search is the design of search space,and the key of convolutional neural network lies in convolution operation.Therefore,in this paper,by observing the characteristics of related computer vision tasks,based on the sensitivity of each task to local micro-patterns,and inspired by the Local Binary Pattern(LBP),this paper designs new convolution types,including Central Pixel Difference Convolution(CPDC),Angular Pixel Difference Convolution(APDC)and Random Pixel Difference Convolution(RPDC),while maintaining the characteristic of normal convolution to be data-driven,it also gives a stronger ability to capture local features.Based on this,the search space of NAS has been redesigned,with four different hierarchical computer vision tasks in image classification(CIFAR10),face perception(RAF-DB,FERET,and RFW),and defect detection(NEU and Magnetic Tile),we introduced neural network search technology into related fields and tasks for the first time in most scenarios,and the network structure obtained by NAS technology search showed high performance and promising results.