Compression And Transplantation Of Object Detection Network Based On Deep Learning

Object detection is an important part of computer vision.It mainly locates the area where a specific object appears in a picture or video and determines the type of object that appears in the image.With the development of deep learning,object detection based on deep learning has made remarkable progress in many fields.However,these deep learning-based object detection algorithms require huge memory space and high-performance computing devices,and cannot be directly transplanted to resource-constrained devices.Therefore,on the basis of summarizing the existing research methods,this paper proposes an improved object detection network and transplants it to the Raspberry Pi for object detection.The main research contents of the paper are as follows:(1)The target detection process is divided into target classification and positioning,and the image quality also determines the accuracy of target classification,but in the actual application process,the image is affected by various external factors,so preprocessing is required before learning the image features.Image preprocessing can eliminate the useless information in the image and retain useful real information,thereby improving the reliability of feature extraction and the classification accuracy of the target to be detected.This paper studies the graying of the image,compares the image denoising effect under different thresholds and threshold function selection,selects the optimal threshold and threshold function for wavelet denoising,and obtains a higher quality image to improve the classification accuracy of the model.(2)In this paper,a lightweight neural network combining extended convolution and depth separable convolution is proposed.Network parameters in order to make up for the current depth is too heavy to in resource-constrained devices and compression problems led to the decrease of the accuracy of the network,proposed the expansion of the convolution and depth of separable convolution of lightweight neural networks,not only greatly reduce the network parameters and calculation of convolution operation,reduce the size of the network,and at the same time of keeping the number of each convolution and invariable,increasing the receptive field,in the process of convolution convolution process to extract more global and the characteristics of the semantic level is higher,to improve the classification accuracy of the network;In addition,by setting the width factor,the three parameters of image resolution and expansion rate are input to further reduce the parameters and calculation amount of the model on the premise of ensuring high accuracy.The experimental results show that the proposed method can effectively compress the network size and obtain a good classification accuracy.(3)This paper aims at the problem that the object detection algorithm SSD(Single Shot Multi Box Detector)uses low-level feature maps to detect objects,resulting in insufficient feature extraction.A multi-scale feature fusion algorithm based on spatial pyramid is proposed to improve the SSD algorithm.Convolution enhances deep feature extraction,and then merges the multi-scale features obtained to enhance the features of low-level feature maps.Experiments show that this method effectively improves the detection accuracy of the network.In addition,the proposed lightweight network is used as the backbone network of the improved SSD,and a object detection network based on the improved SSD network is constructed.Experiments show that on high-performance computers,the network can better balance the detection speed and detection accuracy.The relationship between the two has higher detection accuracy while achieving lower detection speed.(4)In order to apply the object detection network proposed in this paper to embedded and other resource-constrained devices,and can better achieve the object detection task.Based on the Raspberry Pi device,this article uses two methods of Tensor Flow and Tensor Flow Lite to transplant the proposed model and reconstruct the trained model on the Raspberry Pi.Then set the Raspberry Pi camera module to collect images in real time and input each frame of image into the model for target detection.Experimental results show that the network can perform object detection on the Raspberry Pi,and the Tensor Flow Lite framework can effectively improve the detection speed of the model.