Study On The Deep Learning Method For Image Encoding And Object Detection

Object detection is a fundamental and important task of computer vision.It is the basis of many computer vision applications,such as unmanned driving,unmanned monitoring system,character recognition and so on.Currently,Convolutional-NeuralNetwork(CNN)-based algorithms become the mainstream of object detection due to its advanced performance.However,the CNN-based algorithms are limited by their high complexity,and consequently suffer from great computational cost and power consumption when applied to edge devices such as embedded chips.At the same time,edge devices are restricted by the limited bandwidth,so traditional methods also use compressive sensing to compress image data which can reduce transmission bandwidth and power consumption.But limited computation,power consumption,and bandwidth have always been the biggest constraints of the real-time target detection systems on the mobile side.This paper proposed a unified compressive-sensing-enabled object detection deep learning method.It reduces the extra costenabl required for image compression,and promotes the quality of image compression comparing to conventional methods.Finally,this paper implements the proposed algorithm based on hardware and designs an energyefficient edge computing system.The main contributions are as follows:(1)Propose a unified approach for image compression and object detection.The incoherence condition,which is the sufficient condition for recoverable data embedding,is incorporated in the first convolutional layer of the neural network as regularization.It combines the compressive sensing theory and convolutional neural network,enabling the neural network to detect objects and encode images.(2)Propose a new Compressive Convolutional Network(CCN)by improving a classic object detection Convolutional Neural Network.CCN is basically a compressivesensing-enabled convolutional neural network,it optimizes and reuses the convolution operation for recoverable data embedding and image compression instead of calculating object detection and image compression separately.Finally,experiments show CCN improves the efficiency of tradition algorithms by 310% to 500%.On the performance of object detection,CCN have almost no loss.CCN also achieved 3.0dB to 5.2dB higher PSNR for image compression than the examined compressive sensing approaches.(3)This paper implemented an efficient system based on proposed algorithm,and verified the feasibility by using edge computing.CCN shows the ability to build a largescale real-time object detection system by experiments over accuracy,speed and compression rate.