Research On Object Detection Algorithm Based On Knowledge Distillation

At present,Faster R-CNN is one of the mainstream target detection frameworks.The improvement of detection accuracy is accompanied by a large number of parameters and computational overhead,which makes it difficult to apply these algorithms to mobile devices that require storage space and parameters.In order to achieve the goal of balancing detection accuracy and detection speed,a large number of researchers have begun to study methods to compress network models and improve detection speed without reducing the accuracy of network models.Knowledge distillation is one of the results.However,it is very challenging to apply knowledge distillation technology in the field of target detection.The main dilemma currently faced is that after the network model is compressed,the detection accuracy of the model will be greatly reduced.In order to keep the detection performance similar to that of complex networks while reducing the model complexity,this paper proposes an object detection model that combines knowledge distillation and feature fusion.First,the improved knowledge distillation technique replaces the feature extraction network in the Faster R-CNN object detection framework with a shallow network model with fast detection speed and fewer parameters.However,if the improved knowledge distillation method is simply used to train a shallow network to replace the complex network,the detection accuracy of the small network is still quite different from that of the complex network.Therefore,in order to better improve the performance of shallow feature extraction network,feature fusion technology is introduced in the knowledge distillation stage.According to the structure of the residual network,a feature fusion module is designed,so that the output of the complex network can be input to the shallow network in parallel with the input of the shallow network.The main purpose is to enable the shallow network to be fully trained.Experiments on two different datasets show that the detection accuracy of the feature extraction network using the method in this paper is 6.53% and 6.55% higher than that of the feature extraction network without knowledge distillation under the same network size.While ensuring the detection speed is improved,the accuracy of the shallow network after distillation is similar to that of the complex network,which proves the effectiveness of the method in this paper.