Research Of High-throughput Remote Sensing Target Detection And Recognition Based On Lightweight Network

Since the birth of deep learning,target detection technology has been widely concerned and discussed in the field of computer vision.Its application scenarios also involve various fields such as transportation,meteorology,military,education,agriculture and so on.Among them,for some remote sensing image applications,in addition to ensuring the necessary detection accuracy,more emphasis is placed on the real-time information processing,such as traffic situation analysis,positioning navigation,abnormal behavior warning,etc.In this scenario,the timeliness of the target information in the image is very short,and the amount of data obtained in unit time is large,so it is of great significance to inquire fast and effective visual target detection model.In view of the above background,this thesis studies and designs a lightweight convolutional neural network,and further compresses the model to adapt to high-throughput remote sensing image input.The main work includes:1.A lightweight target detection network micro-YOLO is proposed and designed to replace the large Darknet network in the traditional YOLO algorithm.The main idea of YOLO algorithm is used in the network model,that is to predict all kinds of target frames at one time.In the network architecture,the model is compressed by reducing the number of convolution layers and a large number of floating-point operations.In addition,micro-YOLO integrates the core idea of MobileNet,using the depth separable convolution and multi-layer feature fusion algorithm to maintain a good balance between detection accuracy and speed.2.Use a variety of compressing methods to further realize the high-throughput target detection model.Firstly,the network structure is analyzed,and the connection layer with small contribution is cut to reduce the network scale;secondly,store the sparse weight matrix in a compressed way to reduce memory consumption;finally,the weight is quantified,that is to change the number of floating-point operations bit,in order to reduce the consumption of bit width and further reduce the complexity of forward propagation calculation.Finally,through the experimental verification,the proposed micro-YOLO network model is 4.6MB in size.Compared with the YOLO-v2,the micro-YOLO networkmodel gains about 7 times of the speed improvement with the accuracy of 1.64 percentage points,and the model size is about one thirteenth of the YOLO-v2 model.