Research And Implementation Of Lightweighting Algorithms For Target Detection Class Neural Networks

Target detection,a research area of machine vision,has been extensively employed in artificial intelligence,such as intelligent transportation,medical care,and wearable devices.However,since high-performance GPUs are often not available in these scenarios,how to apply complex network models to low-power,low-performance devices has become a worthwhile research direction.Beginning with an examination of the features of the target detection network,this thesis examines two approaches to lightweighting the model: knowledge distillation and sparse pruning.Subsequently,experiments were conducted to demonstrate the effectiveness of the strategy proposed in this thesis.Finally,this thesis proposes a complete lightweight processing and deployment scheme,which can reduce the model size and improve its inference speed while ensuring the model accuracy,and designs a target detection model lightweight system with this scheme as the core algorithm.The main contents of this thesis are as follows:1.Propose a knowledge distillation method for target detection networks.Firstly,the defects of traditional knowledge distillation algorithms are analyzed,their distillation losses are analyzed and recombined,and a new classification distillation loss is proposed,followed by the distillation of the prediction frame of the target detection network combined with the characteristics of the target detection network.The final knowledge distillation scheme of this thesis is proposed by combining the improved methods of classification distillation and localization distillation.2.Put forward a model lightweighting and deployment scheme based on pruning algorithm.To begin,an examination of existing pruning techniques is conducted,followed by the suggestion of this thesis’ s channel pruning system.Subsequently,the superiority of this thesis’ s system is demonstrated through comparison experiments.This thesis proposes a lightweighting model by combining knowledge distillation and channel pruning methods,which is then implemented on the low-power platform Jetson Nano.Experiments demonstrate the efficacy of the algorithm and the correctness of the combination approach.3.The target detection model lightweighting system is designed and implemented.Combined with the knowledge distillation and channel pruning methods studied in previous,the outline design and detailed design are carried out according to the requirement analysis,and the target detection model lightweighting system is coded and implemented according to these designs.Finally,the application feasibility of the lightweight processing scheme of the target detection model proposed in this thesis is verified by testing.