Research And Implementation Of Tiny Object Detection Algorithm Via Generative Adversarial Network

Object detection is one of the most basic and challenging problems in the field of computer vision,which is the basis of solving complex or high-level vision problems.The task of object detection is to find target instances from a large number of predetermined categories of natural images.At present,object detection has a very wide range of applications,including robot vision,automatic driving,security and augmented reality,and object detection on the UAV platform is one of the hot research fields.The advantages of UAV are small,flexible,maneuverable and easy to deploy.Combined with object detection methods,UAVs have important application value in disaster rescue,environmental protection monitoring,traffic monitoring and battlefield investigation.However,due to the small size of targets,the complexity of background,the illumination and camera jitter,the object detection based on UAV is difficult.At present,many researches focus on solving the problem of tiny object detection.SSD uses shallow feature mapping to represent small targets.However,the semantic information of shallow features is weak and the distinction is poor,which leads to the emergence of false positives.In addition,FPN uses feature pyramids to represent objects of different scales.Deep feature maps with strong semantic information are sampled up and fused with shallow features,however up-sampling may produce artifacts and therefore reduce detection performance.Many state-of-the-art detection algorithms have poor detection effect on tiny targets.Therefore,it is of great significance to study tiny object detection algorithms with high-precision and high-performance in complex scenes of UAV.In order to solve the above problems,this paper proposes a new two-stage small target detection model based on the generative adversarial networks(GAN).In the learning process,Gan makes full use of the structural correlation between objects of different scales,enhances the similar representation of small target and large target,and makes small targets as easy to be detected as large targets.The process of object detection is divided into two stages.In the first stage,the input image is extracted by the deep convolution neural network,and the feature of image is enhanced by the super-resolution network.The features extracted by the two methods are fused by the way of element summation to get the preliminary extracted features.By using the region proposed network,the possible position of the target and the positive and negative samples are preliminarily extracted.In the second stage,we judge the exact location and specific category of the preliminary extracted proposed regions,which is used as the final detection result of our proposed network.The model is tested on the public standard data set PASCAL VOC and MS COCO.The results show that the model has better super-resolution reconstruction ability and detection ability for small and medium targets in natural image.In addition,the model is deployed to the UAV platform,which can effectively locate the precise position of the ground vehicle in the complex air environment,reflecting the good practicability and robustness of the model.