Research And Implementation Of Remote Sensing Image Object Detection Method Based On Lightweight Network

With the rapid development of remote sensing information technology and deep learning technology,the performance of object detection in visible remote sensing images has been significantly improved and has played an important role in land surveying,urban planning and vital project construction fields.However,visible remote sensing images usually suffered in complicated background,object scale diversification,and small objects,which makes it more difficult to detect the object in an optical remote sensing image.In recent years,although the deep learning-based remote sensing image object detection algorithms have higher accuracy,these models are complex with large parameters and high calculation,and the detection performance of small targets still has a wide space for improvement.Therefore,the research on the object detection algorithm of remote sensing image with better and more efficient has important theoretical significance and application value.This paper conducts the research on the basis of the YOLOv3(You Only Look Once V3)real-time object detection algorithm to improve the performance in visible remote sensing images.The improved YOLOv3 has been light-weighted to achieve more efficient and high-precision object detection.A remote sensing image object detection system is designed and developed,which is convenient for users to complete the task of object detection through the interface with pictures and texts.The main work of this thesis are as follows:(1)In view of the complex background,object scale diversification and small targets in visible remote sensing images,this thesis proposes an improved object detection method based on YOLOv3 algorithm.Firstly,based on the YOLOv3 model,the feature map detection structure is improved to introduce shallow features with rich location information into the feature detection map to enhance the detection ability of the model algorithm for small targets.Then,the anchor frame matching ability is improved by K-means framing mechanism.The RFB(Receptive Field Block)is introduced into the feature detection layer to increase the receptive field of the feature map while fusing multi-scale feature information to further improve the detection ability of the model.The effectiveness of proposed method is verified through comparative experiments on the remote sensing image dataset and its background mask dataset.(2)The improved model has a relatively complex structure,a large amount of parameters and calculations,and a long model loading time,which makes it difficult to meet the needs of real-time object detection.Therefore,the proposed model is compressed to obtain a lightweight object detection algorithm.The compression process consists of three main steps.Firstly,a sparse training is used to test the importance of channels in every convolutional layer.Next,the channels with low importance are pruned by different channel pruning rates,and the corresponding convolution kernels are cleared according to the channel pruning results.So that the coupling relationship between convolution layers is taken into account while implementing channel pruning.Finally,the lightweight network model is fine-tuned to evaluate and restore its performance of object detection.(3)A remote sensing image target detection system is designed and developed by using Python language,B/S(Browser/Server)structure,Flask and Vue frameworks.The system mainly includes user information management,model management,detection management and page management,etc.The functions of the system are tested through the experimental datasets,object detection models and related configuration files.It has realized the rapid detection of remote sensing image objects,which can be used as a reference for the construction of large-scale object detection system.