| As an important task in computer vision,object detection is the basis of semantic segmentation,scene understanding and other visual tasks.The target detection model is usually deployed in the cloud and edge to extend its application.It is of great significance to use the neural network architecture search technology to search the target detection model that can improve the accuracy without increasing the amount of computation and parameter number.In this thesis,the existing target detection model and heavy parameter technology are combined.In the training stage,the reparameter branch expansion is carried out for the convolution operation whose convolution kernel size is greater than 1 in the main network of target detection.In the case of reparameter branch expansion,the neural network architecture search method is used to efficiently search and evaluate the performance benefits of different branch combinations.The specific content includes analyzing the existing main target detection network,screening the most commonly used Res Net network and lightweight Det NAS network for improvement,designing the reparameter search space and heavy parameter branch selection algorithm to complete the automatic selection of branch combination and model building;In order to make the width and depth of the network model adapt to the reparameter branching structure,a model morphism algorithm is designed to adjust it.The low fidelity performance evaluation strategy was used to sort the model performance,and the Pareto front algorithm was selected to screen out the optimal model under the three metrics of accuracy,parameter number and computational amount.After the optimal model is trained,the reparameter branches are fused into the original convolution operation in the inference stage to reduce the computation and parameter number of the inference model.Compared with Res Net50,the Top1 accuracy of Rep Res Net classification model based on Image Net data set is improved by 1.4.The target detection model Rep Res Det optimized on MS COCO data set can increase the mean accuracy by 1.55 under the premise that the number of parameters of relatively advanced FPN is basically unchanged.When the mean accuracy of Rep Det NAS is only 0.23 lower than that of FPN,the number of parameters decreases by 53.5%.The model deployment tool Tensor RT is used to accelerate the Rep Res Det inference,and the average recognition speed of a single image can reach 0.072 s.When the recognition model is deployed on Jetson TX2,an artificial intelligence embedded platform,the average inference speed can reach 0.32 s per frame.The target detection model designed in this thesis has the characteristics of low parameter number,computation amount and high precision.It has potential application value in product quality inspection,security and other fields. |
