Research On Lightweight Faster-RCNN Algorithm Based On Knowledge Distillation

In recent years,as a key technology in face recognition,automatic driving,industrial detection and other fields,object detection has attracted more and more scholars’ attention.With the rapid development of deep learning technology,the target detection network has also changed from traditional manual design to a deeper network framework,which has not only improved the speed but also brought a qualitative leap in detection accuracy.However,the more complex the deep learning algorithm framework is,the more resources and memory it takes up.The complex network structure and expensive running costs also lead to many algorithms being unable to be applied in real life.As an algorithm with high detection accuracy,Faster-RCNN has a complex network structure and parameters,which is difficult to be applied in practice.To solve this problem,this paper aims to lightweight Faster-RCNN algorithm.The main research of this paper is as follows:(1)Faster-RCNN algorithm model based on RV-C(Rep VGG-CBAM).After the algorithm is lightened,the accuracy is reduced,so the feature extraction network of the algorithm is processed before the lightening.The Faster-RCNN algorithm uses VGG16 as the feature extraction network,which has deep layers and contains many convolution layers and pooling layers.However,part of the feature information in the image will be lost in the process of convolution and pooling.To solve this problem,Rep VGG feature extraction network is used in this paper to merge with the Faster-RCNN algorithm.The Rep VGG network contains a residual module to reduce the convolution operation and improve the effect of feature extraction.In addition,the attention mechanism is embedded in the feature extraction network to obtain important features in the image and process the input image from both spatial and channel dimensions,so as to improve the detection accuracy of the Speed-RCNN algorithm for small targets and the average detection accuracy of the algorithm.(2)Faster-RCNN Algorithm model with D-RV-C(Distillation-Rep VGG-CBAM).After feature extraction,the single-stage target detection algorithm directly sends the image to the subsequent network for target classification and regression,while FasterRCNN is a two-stage target detection algorithm.After feature map extraction by feature extraction network,the feature image needs to be sent to the RPN part of the algorithm to get the candidate box of the target,so as to improve the positioning accuracy of the target.However,RPN has a large number of parameters,so the detection rate of the algorithm is affected to a certain extent.Therefore,this paper carries out model compression processing for the Faster-RCNN algorithm,and adopts the knowledge distillation method to carry out distillation processing for each part of the algorithm to reduce the model size and improve the detection speed of the algorithm.The improved algorithm was tested on PASCAL VOC 2007,PASCAL VOC 2012 and COCO datasets.Compared with the original Faster RCNN algorithm,the model size after knowledge distillation was reduced by 1/3,and the average detection accuracy was only reduced by 0.7%.Compared with other Faster-RCNN algorithms improved by knowledge distillation,the detection accuracy was also improved.Figure [33] Table [13] Reference [71]...