Lightweight Object Detection Algorithm Based On Deep Learning

In recent years,deep neural network has achieved great progress in the field of computer vision.However,these deep learning-based object detection networks often suffer from large model parameters and excessive memory consumption,which have high requirements for hardware devices and resources.The application of object detection algorithm has great challenges,especially in the embedded devices with limited computing and storage resources.Therefore,the main research in this work lies in investigating lightweight deep networks for the application of object detection.Aiming at the object detection of classic network SSD?Single Shot multibox Detector?and RFBNet?Receptive Field Block Network?,this thesis proposes two structural pruning schemes.The generated lightweight object detection network can keep good accuracy and reduce the size of network model,at the same time,it also improves the detection speed.The completed work mainly includes:?1?An object detection network pruning scheme with the fusion of knowledge distillation is proposed.This scheme is based on L₁ norm as the importance evaluation criterion of each filter in the convolutional layer.For the object detection network SSD,this scheme adopts single layer pruning scheme to pruning a single convolutional layer of the network at first.Then considering the time cost of iterative pruning for the entire network with a single-layer pruning scheme,multi-layer pruning scheme is introduced to achieve pruning of the entire network at once.Finally aiming at the precision loss of multi-layer pruning SSD,this scheme further introduce an improved version of the multi-layer pruning scheme by integrating knowledge distillation.The original network can be used to guide the pruning network to carry out retraining so that the pruning network could achieve better detection performance.The experimental results show that the object detection network pruning scheme can effectively reduce the size of the network model while maintaining the network detection accuracy.With the increase of pruning rate,the number of parameters and computation of the model will be greatly reduced.Compared with the network without pruning,the generated lightweight object detection network has a faster detection speed.?2?An object detection network pruning scheme based on scaling factor is proposed.The scaling factor of the batch normalization layer is used as the channel importance factor in this scheme.Aiming at the object detection network RFBNet,the local pruning scheme,global pruning scheme and double criteria pruning scheme are designed.In these schemes,the local pruning scheme divides the original RFBNet into two parts,and each part can be pruned separately,or two-stage pruning scheme can be carried out.The global pruning scheme can evaluate the importance of each filter in the convolutional layer in the whole network and prune the entire network at once.Moreover,we propose double criterion pruning scheme based on the presented global pruning scheme,which can evaluate the importance of the filter in the whole network and its convolutional layer.The experimental results show that the object detection network pruning scheme based on scaling factor can effectively reduce the size of the network model and generate a lightweight object detection model while ensuring the network detection performance.