Vehicle Detection Based On Deep Learning In Remote Sensing Images

As one of the important parts of remote sensing image interpretation,vehicle detection has a wide range of application prospects in areas such as smart cities,autonomous driving and battlefield situational awareness.Vehicle targets are more difficult to detect than other ground targets because of some image collection difficulties lead to incomplete vehicle datasets,the small differences between the vehicle and the background,the different scenes in which the vehicle targets are to be detected,the different resolutions of the vehicle images,the variety of vehicles and the low illumination at night.Deep learning models provide a new solution for vehicle target detection,but the performance of detection of weak,mixed,dark and new vehicle targets still needs to be improved.Therefore,it is of great theoretical significance and application value to investigate vehicle target detection methods under incomplete sample conditions,low signal-to-noise ratio,blending and low illumination.Using a deep learning framework,the paper proposes four works on data augmentation,image enhancement,two-stage detection models and multi-modal image fusion to address the difficulties of angular differences and grey inversion in incomplete data sets,low signalto-noise,blending and low illumination in vehicle target detection,as follows:A data augmentation scheme based on a priori knowledge of the scale and grey distribution of the test set,the scale augmentation scheme and the grey augmentation scheme,is proposed to solve the problem of incomplete data sets due to the difficulty of collecting infrared image data of vehicle targets.The scale augmentation scheme augments the data by scale scaling,and designs a three-branch detection model that includes the original data branch,the scale-broadening and scale-down branches,and adds an interbranch Generative Adversarial Network mapping module to maintain the consistency of the features extracted from the three branches.The results of testing a large number of angular difference datasets show that the scale priori-based data augmentation scheme can effectively improve the detection performance of oblique vehicle targets.The grey augmentation scheme augments the data with grey-scale inversion,and designs a twobranch model containing the original data branch and the grey-scale inversion branch,with the addition of a Generative Adversarial Network mapping module generated between the branches.The results of testing a large number of grey-scale difference datasets show that the grey priori-based data augmentation scheme can effectively improve the performance of vehicle target detection with large grey variations.A vehicle target detection method based on target enhancement by cycle Generative Adversarial Networks is proposed to effectively solve the problem that vehicle targets with low signal-to-noise ratio are difficult to detect accurately by enhancing vehicle target features in satellite images.Aiming at the problems of low contrast between vehicle and background in visible satellite images and inadequate feature representation by conventional feature extraction algorithms,this paper uses cycle Generative Adversarial Networks to enhance the features of vehicle targets,trains low-quality to high-quality image generation models using unpaired satellite images and aerial images,and introduces detection supervision branches to enrich the features of vehicle targets in the generated images.The experimental results show that the vehicle detection method based on target enhancement effectively improves the performance of low signal-to-noise ratio vehicle target detection in satellite images.A two-stage vehicle target detection model based on the full convolutional detection model is proposed to effectively solve the problem of poor detection capability of existing detection models due to the large variation of vehicle target scenes,sizes and types.To address the problem of difficult detection of mixed vehicle targets with multiple scenes,scales and types,this paper designs a candidate frame extraction module,a two-stage positive and negative sample sampling mechanism module and a two-stage classification module to improve vehicle target detection performance.The candidate frame extraction module uses full convolutional detection model to generate candidate frames for vehicle targets covering multiple scenes,types and scales.The two-stage positive and negative sample sampling module improves the sampling of small-sized vehicle targets.The twostage classification module optimises the candidate frame classification.The experiments demonstrate that the two-stage detection model has good detection effect on mixed vehicle target detection.A vehicle target detection algorithm,which fuses visible and infrared images through Generative Adversarial Networks,is proposed to effectively solve the problem of vehicle feature loss in visible images under low illumination and improve the performance of vehicle target detection under low illumination.To address the problem of insufficient vehicle features in low illumination,this paper uses visible and infrared branches based on generative adversarial networks and an attention fusion module.The visible and infrared branches use an image generation process to fuse image features and target features respectively.The attention fusion module fuses the features of the two branches by means of an attention mechanism.The experimental results validate the effectiveness of the fusion model for low-light vehicle target detection.