Research And Application Of Lightweight Objective Detection Arithmetic

With the development of deep learning,objective detection technology has made a major breakthrough,and many excellent convolutional neural network models have emerged.The calculation and “volume” of these models are big,it can only on high-performance servers.However,there is a huge demand for object detection on terminal mobile devices,which is limited by the computing resources and battery capacity.It is required that the object detection model needs to be small enough and fast enough and stable enough to meet the detection accuracy.Therefore,based on the research of existing object detection models,this thesis will carry out in-depth research on the stability,lightweight design and improvement of detection accuracy of the model,and propose a feasible design idea of lightweight object detection model.Last but not least,this lightweight model will be deployed on mobile terminals such as mobile phones.Firstly,the problem of model instability caused by the use of different scaling algorithms in the image preprocessing stage of the lightweight model is studied.Due to the different calculation database on different mobile terminals,the scaling algorithm of the model is difficult to unify.If the image scaling algorithm is fixed during preprocessing,the model will be "over-fitting" into the zoom mode,and the adaptability to other scaling algorithms is poor,which is disadvantageous for the lightweight model training.To this end,this paper proposes that a variety of image scaling "modes" should be enhanced in the training data to improve the diversity of training data.Training the neural network to enhance its generalization ability to the different scaling algorithms by combining the knowledge distillation technology.Experimental results confirm the effectiveness of the method.Secondly,this paper studies the lightweight object detection model based on Retina Net,and improves the model for both the lightweight design and the detection accuracy:1)The three components of Retina Net(Backbone Network,FPN,and Detection Head)are analyzed separately for the impact on the detection accuracy after replacing it with a lightweight network structure.This paper finds that the FPN part is least sensitive to the lightweight network design,and the branch of the output object classification in the detection head is more sensitive than the branch of the output object position.Therefore,this paper proposes an improved network block based on Mobile Net V2 to lighten the FPN part of the Retina Net model,to replace the FPN with a slight loss of accuracy,and to reduce the calculation by reducing the number of anchors on the bottommost feature of the FPN;2)This paper proposes a knowledge distillation method based on detection frame information to improve the accuracy of the detection model.Based on the traditional knowledge distillation,this paper outputs the additional detection frame IOU confidence in Retina Net,and uses this confidence to distill the detection frame information to reduce the precision loss caused by the weight reduction of the detection model.Finally yet importantly,this paper constructs a practical clothing detection dataset in the academic and industrial fields.Combined with the object detection technology proposed above,a lightweight clothing detection model is designed and implemented.The deployment on the mobile phone is completed by replacing the FPN upsampling algorithm,BN fusion,parallel NMS etc..The lightweight clothing detection model has been applied into a large Internet company,which not only reduces the consumption of enterprise network bandwidth,but also obtains an average of millions of visit traffic per day due to its high detected efficiency.