Design And Implementation Of Lightweight Target Detection Network Based On Perfoemance Predictors And Search

Object detection is one of the important applications of deep learning algorithms,which has achieved far better performance than traditional detection algorithms in scenarios such as vehicle detection and face detection.However,deep learning-based convolutional neural network algorithms have large complexity characteristics and parameter number scale,and their performance evaluation takes a large amount of time.How to efficiently design lightweight neural networks that meet the requirements of hardware computing power resources is a difficult problem facing the perception algorithm based on edge devices.Therefore,this thesis uses a neural network search method based on neural network performance predictors to achieve the design of lightweight object detection networks.A performance predictor based on Gaussian process modeling of neural network structure and accuracy is designed,and a progressive neural network search algorithm based on the predictor is proposed.In the search process,a reward function balancing model performance and hardware constraints was introduced to obtain a lightweight backbone network that meets the hardware characteristics,and ultimately a lightweight object detection algorithm is obtained.Finally,the effectiveness of the above model and algorithm is verified based on the Xilinx ZCU102 hardware platform.The main research contents are as follows:1.In response to the problem of huge computational and time costs caused by the need to train and evaluate a large number of network structures in neural network search algorithms,this thesis designs a neural network performance predictor based on Gaussian processes.To address the issue of feature representation in direct-link neural networks and the insensitivity of Gaussian processes to positional information encoding,new encoding methods and kernel functions are proposed to improve the fitting ability of the performance predictor.The designed predictor achieves a Kendalltau coefficient of 91.3% and a Spearman coefficient of 99.0% on the OFA-Proxylessnet dataset,and achieves the Top-10 ranking in the CVPR22-NAS Performance Estimation Track competition.2.Based on the neural network performance predictor,this thesis proposes a multi-task progressive search method to address the problems of how to further improve the fitting ability of the performance predictor to a huge search space with limited samples and how to incorporate hardware characteristics into algorithm design.The method uses the predictor to sample network structures with good performance,gradually refines the sampling space sorting,and improves the predictor’s prediction performance for top-level network structures.In addition,a reward function based on Pareto optimal solutions is proposed to balance model performance and hardware constraints during the search process,fully considering the hardware characteristics of the deployment terminal.The final searched PRDNet achieves the Image Net Top-1accuracy of 76.4% and the speed of 244.36 FPS on the ZCU102,realizing the expected design of a lightweight backbone network.3.Based on the efficient lightweight backbone network discovered through the search,the object detection network is trained using reparameterization method,and achieved an m AP@0.5 of 81.5% on the VOC2007 test set.Then the object detection network is quantized,compiled,and deployed using the Vitis AI compilation tool.Based on the Xilinx ZCU102 platform,the network achieves a detection frame rate of35.5 FPS,enabling efficient operation of the perception system and validating the effectiveness of the proposed method.