Specific Object Detection And Recognition In Optical Remote Sensing Images

Object detection in optical remote sensing images has always been a hot spot in the field of remote sensing applications.With the rapid development of image sensor technology and aerospace industry,the resolution of remote sensing image is getting higher,and the resolution of the optical remote sensing images is as high as 0.3m per pixel.In such high-resolution optical remote sensing images,the contours and the textures of objects are clearly observed and informative for object detection,and object architectures are closer to characteristics of human ocular perception than SAR images.In a word,the study on the object detection in optical remote sensing images has great significance in aspects of theoretical research and practical application.The topic of this article is derived from the "Multi-angle optical image registration and object change detection technology research" project,which uses ten typical man-made targets such as airplane,ship,storage tank,harbor,bridge,and vehicle as objects of detection.Specific object detection algorithms for optical remote sensing image are studied to meet the application needs of more recognition accuracy and higher speed.The main work of this paper is as follows:(1)Applying the Faster R-CNN algorithm to the object detection task,adaption changes are applied at the application level targeted at specific requirements of optical remote sensing images.For the problems existing in optical remote sensing image of color-texture interference,rotation-scale variance and shape-similar interference,eight enhancement processing for training data are performed,such as flipping,panning,rotation,and color jitter,etc.It is proved that the increase of the number of samples and transfer learning significantly enhances the training results of the model,and realized the detection and recognition of remote sensing multi-category man-made objects with an average accuracy of 71.2%.(2)The YOLO v2 algorithm is applied to the object detection task of optical remote sensing imagery.The joint training mechanism is used to train the classification network at first using the remote sensing classification dataset and the network achieves a top-5 accuracy of 88.2%,and then use the weight parameters of the first 23 layers to train the detector,which can predict the position of multiple bounding boxes and category simultaneously.The recognition speed of the detector can reach 67 FPS with a detection time of about 15ms/image,and meet the real-time detection requirements.The combination of classification data with larger amount and detection data with better label greatly reduces the training cost and expands the scope of detection model.