Research On Object Detection Network Lightweight Method With Constrained Resources

Deep learning technology has advanced significantly in the area of computer vision during the past several years.However,in terms of computational and storage resources,the problem of higher redundancy in the detection network poses a significant obstacle to the project’s large-scale deployment of application object detection techniques.Hence,for object detection algorithms to be widely used in current engineering equipment,Research on object detection network lightweight methods with constrained resources is essential.This paper provides a collection of lightweight algorithms for object detection networks in the context of a full-process survey.These strategies lower the object detection network’s resource consumption and increase the practicality of the object detection technology.The following three elements are part of the main work:The present object identification methods are sophisticated,and their efficiency in resource-constrained circumstances is not completely considered in the design.This paper proposes a lightweight object detection network based on a low-cost convolution module to address this issue.First and foremost,a more efficient low-cost convolution module CIRBlock is developed that is more suited for application in lightweight networks than standard convolution.Then,CIRBlock is used to construct CIRNet,a lightweight feature extraction network that is better suited for use on lightweight detection networks than standard convolution.Finally,construct the lightweight object detection network CIR-YOLO by combining CIRNet and lightweight object detection components developed with CIRBlock.Experiments show that the lightweight object detection network CIR-YOLO,which is gradually built from the redesign of the lightweight convolutional module,has a 1.9% higher m AP than YOLOv5-s under similar resource occupation.It can better balance resource consumption and detection accuracy than the existing lightweight object detection network.Given that knowledge distillation algorithms can improve the accuracy of lightweight object detection networks by utilizing complex and high-precision detection networks,this paper proposes a more effective knowledge distillation algorithm for object detection networks based on augmented instances.Firstly,the knowledge distillation algorithm in this paper uses the instance-based object detection network as the fundamental framework after examining its advantages and disadvantages.Then,a great select instance module is created,which addresses the issues of different output sizes and inaccurate calculation of differences between instances of teachers and students.This module combines the output characteristics of the object detection network.Finally,a variety of knowledge based on features,responses,and relationships is constructed when combined with the lightweight object detection network designed in this paper.They can ensure that the student network can mine more potential information from the teacher network,improving the lightweight detection network’s accuracy.Experiments show that the object detection network’s knowledge distillation algorithm proposed in this paper can improve the m AP of the CIR-YOLO by 0.9% while keeping the network structure unchanged.A lightweight object detection network compression algorithm based on cooperative channel pruning and tensor decomposition is proposed to address the problem that the lightweight object detection network after training still has more weight redundancy.The first step is to mathematically unite the channel pruning and tensor decomposition.And a network cooperative compression sensitivity analysis algorithm based on the greedy approach is developed.Then,the network weight is gradually compressed layer by layer based on the gradient descent algorithm and Markov decision process,ensuring that the impact of a single compression operation on network performance is as minimal as possible.The results of the final experiment demonstrate that the proposed cooperative channel pruning and tensor decomposition object detection network compression algorithm can fully utilize the sparsity and low rank of neural networks to compress the detection network and enhance the operational efficiency of the detection network in resource-constrained scenarios.