Research On The Object Detection Model Lightweight Algorithm Based On Convolutional Neural Network

Object detection is one of the basic problems of computer vision.It is widely used in many fields,such as intelligent transportation,intelligent security and industrial defect detection.The object detection algorithm based on deep convolution neural network shows excellent performance in different scenarios,but the huge demand for computing and storage resources seriously limits the landing application of relevant algorithms on edge devices with limited resources.This thesis focuses on the lightweight algorithm of object detection model based on deep convolution neural network.From the model acceleration and performance improvement,the channel pruning algorithm derived from dynamic gating system and context aware knowledge distillation algorithm suitable for object detection model are proposed respectively.Using their lightweight characteristics,a lightweight joint algorithm framework is designed.The specific research contents are as follows:1.Aiming at the problem that the performance of traditional model pruning algorithm decreases sharply due to incomplete evaluation of pruning objects,a channel pruning algorithm derived from dynamic gating system is proposed on the basis of introducing dynamic channel pruning strategy to provide appropriate evaluation criteria for network nodes.This method fully integrates the input oriented dynamic adjustment advantages of dynamic pruning and the model simplification ability of static pruning,and generates an object detection model with simpler structure and faster speed on the premise of ensuring accuracy.The experimental results on typical object detection VOC dataset fully prove the effectiveness and universality of the algorithm.The model size of different models is reduced by more than 39.6%,the forward reasoning speed is reduced by more than 42.9%,and the performance is reduced by no more than 1.6%.2.Aiming at the problem that the accuracy of simple object detection model is always low,a context aware knowledge distillation algorithm suitable for object detection is proposed in this thesis.The algorithm includes the context awareness module and self-attention mechanism,making full use of the context information of the detected object,the refined teacher network knowledge is extracted from the spatial domain and channel domain,and the refined knowledge is used as additional supervision information in the process of student network training.In the validation experiments of the effectiveness and generality of the algorithm,the detectors based on different backbone have achieved more than 2.9% performance improvement on VOC and KITTI datasets.3.Aiming at the problem that the object detection model based on convolutional neural network has a large number of model parameters and slow forward reasoning speed,which leads to the difficulty of model deployment.A lightweight joint algorithm based on model pruning and knowledge distillation is proposed in this thesis.Firstly,the channel pruning algorithm derived from the dynamic gating system is used to prune to reduce the number of parameters.Then,the context aware knowledge distillation algorithm is used to improve the accuracy,and a lightweight model with small parameters,fast reasoning speed and high accuracy is obtained.While maintaining the accuracy of the baseline model,the parameter compression exceeds 48.2% and the amount of calculation decreases by more than 53.1%,which proves the effectiveness of the proposed algorithm framework.