Research On Object Detection In Remote Sensing Image

Remote sensing images can be widely used in urban planning,environmental governance,agricultural production,ship identification,and maritime monitoring due to their high resolution and rich spectral information.This paper studies the object detection technology in remote sensing images.The main results of the paper are:1.Aiming at the shortcomings of synthetic aperture radar(S AR)image information such as high noise,great influence by interference,and less color information than optical image,a space-frequency transform algorithm for single-channel SAR image is proposed,and two fusion networks of channel fusion and random selection are constructed.Experiments on some public SAR datasets show that the performance of the fusion network is better than that of the ordinary network,and the performance of the randomly selected fusion network is no less than that of the channel fusion network without additional parameters for inference,which proves the feasibility of the datacentric idea.2.A new RDM matching algorithm and an AE coding algorithm are proposed respectively.The RDM algorithm can improve the matching effect of objects with large aspect ratios by discarding geometric information and adopting the numerical representation information of the rotating frame,the AE algorithm can Solve the problem of nonlinear relationship between the intersection ratio of two rotation boxes in the original encoding algorithm,thereby improving the performance of the network.The experimental results show that the RDM and AE algorithms can be embedded to most of the anchor-based object detection network without additional network parameters,and are applicable to both single-stage and multi-stage detection networks.3.Combined with model pruning and neural architecture search technology,a lightweight and accelerated scheme for remote sensing object detection model is proposed.Different model light-weighting and acceleration strategies corresponding to different model production environments are proposed,which can improve the inference speed of the network while guaranteeing the accuracy of the results.