Multi-scale Dense Object Detection For Optical Remote Sensing Images

With the continuous development of remote sensing information technology and visual intelligence technology,remote sensing object detection technology continues to provide services in many fields.Researchers at home and abroad have done a lot of research on remote sensing object detection methods based on traditional feature extraction methods and deep learning.However,the existing object detection methods still have the following deficiencies:(1)The traditional object detection methods have relatively shallow feature representation capability and the candidate region generation algorithm has relatively low computational performance.(2)The existing deep learning detection algorithm has weak generalization ability and is only suitable for specific object types.It is difficult to solve the challenges of different object scales and dense aggregation in optical remote sensing images.Therefore,the current optical remote sensing image object detection algorithms still have broad development space,and the research on more efficient detection algorithms has important academic and application value.This thesis analyzes the shortcomings of current object detection methods and theories,studies the detection technology of remote sensing objects from any angle,designs a object detection network suitable for various scales and densities,and tests various methods to enhance the spatial transformation capability of the detection network.The specific work content is as follows:(1)Aiming at the difficulty that there are many false and missed detections in dense object detection of optical remote sensing images,through analyzing the characteristics of dense objects,a joint branch detection framework RA-CNN is designed,in which parallel vertical boxes and quadrangle boxes branches jointly participate in the optimization of network weights.An adaptive ROI feature extraction method for candidate regions is designed to achieve a balance between background noise suppression and spatial information preservation,and ROI context information is obtained through a global large convolution module.For the problem of sample imbalance in dense objects and the inconsistency between Smooth L1 loss and IOU evaluation,an improved loss function is proposed: Gaussian modulation classification loss and joint rectified location loss.Through these two losses,multi-task learning is carried out and deformable convolution is adopted in the basic network structure to enhance the spatial transformation capability of the model,so that the classification and positioning accuracy of dense objects and small objects are significantly improved.(2)Aiming at the difficulty of low detection performance of small targets in multi-scale object detection of optical remote sensing images,by analyzing the influence of different scale features on detection accuracy,a multi-receptive field feature fusion module RF-Inception based on atrous convolution and Inception structure is designed,and the fused features are automatically screened by residual module Res CSA which combines channel attention and position space attention.In the detection output part of the network,the cascade structure can reduce the information loss of small objects and improve the overall performance under different IOU thresholds.A comprehensive comparative experiment was carried out on two remote sensing data sets by combining the optimized detection methods for dense and multi-scale objects.The experimental results show that the proposed network structure and key algorithms have high accuracy and good generalization performance for multi-scale and dense object detection in optical remote sensing images.(3)According to the research results of optical remote sensing image object detection in this thesis,a high-resolution optical remote sensing image object detection application software based on web platform is designed.The organization structure of the software,the modular interface of the object detection model and the scheduling process are described in detail.Finally,the high-efficiency detection of multiple objects with different resolutions,scales and densities is realized.