Lightweight Detection Algorithm Based On Transfer Learning

With detection accuracy that surpasses traditional algorithms,the algorithms based on deep learning have gradually become the mainstream algorithms for target detection tasks and been widely used in robots.However,this type of algorithm relies on a large amount of data annotations in the model training process and also has great requirements for the computing power of the hardware platform,which limits its application.With the ICRA RMUA artificial intelligence challenge as the background,this paper studies the automatic labeling of training data in target detection tasks,the domain adaptive training of the model,and the lightweight of the model.The main research results of this article are summarized as follows:First of all,the camera model is established,the coordinate system definition and robot position analysis are given;The task requirements and indicators of visual inspection in robot competitions are analyzed and the sensor scheme of the visual inspection system is given;Secondly,taking into account the limited computing power of the robot platform and the high-frequency,low-latency detection index requirements,a target detection framework with global detection plus ROI tracking is proposed and the detection and tracking network is designed based on YOLOv3;Feature extraction network are designed based on Mobile Net V2 and YOLOv3-tiny and improved the loss function of YOLOv3 to reduce the requirements for the balance of training samples.Thirdly,a method of automatic labeling of training data is studied.A visual simulation environment is built based on UE4,which can simulate the competition environment very realistically,and can render virtual pictures with target masks;An automatic labeling method for training data based on the visual simulation environment is designed;In order to reduce the difference between virtual pictures and real pictures,data preprocessing methods are proposed.Fourthly,in order to reduce the loss of model migration caused by the difference between simulated data and real data,a network model domain adaptive training method based on adversarial features is designed which aligns the data distribution at the feature extraction level.And consistency regularization method is introduced to improve the robustness of the model.Then,the model lightweight method is studied.A model pruning method based on Gumbel Softmax sampling is proposed.The Gate Module module is designed for training a sparse network model with a dynamic pruning strategy,and parameters pruning mothed for the sparse model is applied to achieve further compression of the neural network.Finally,the proposed lightweight detection algorithm is implemented and deployed.Experimental results show that the use of domain adaptive methods and model pruning methods,completely using simulation data training target detection model,have reached the index requirements of high frequency,low latency and high accuracy in different test environments,which verified that simulated data sets can be used as an alternative of artificially labeled data for training models.