Visible Spectral Remote Sensing Image Classification And Detection Using Deep Convolutional Networks

With the increasing resolution of visible satellite remote sensing and aerial remote sensing images,more valuable information can be interpreted from these images.The classification and detection of remote sensing images is an important procedure for intelligent interpretation of remote sensing images,and it is also a research hotspot in this field.The traditional methods are usually adopted handcrafted features and shallow classifiers such as regression and SVM.However,due to the complex background of visible remote sensing images,large intra-class scene differences,and large changes in target scale,the traditional methods encountered a bottleneck in performance improvement.The deep convolutional network provides a new effective method for the classification and detection of remote sensing images through the multi-layer abstraction mechanism and the autonomous learning of high-level essential features.Therefore,this paper mainly studies remote sensing image classification and detection algorithms based on deep convolutional networks.The specific contents are as follows:(1)A remote sensing image classification algorithm based on random multi-selective residual network is proposed.The deep convolutional network model has been successfully applied in image classification,object detection and other issues.However,a single type of network still cannot obtain good classification results for scenarios with large intra-class differences and large scale changes.It is beneficial to improve classification accuracy by integrating multiple networks for classification through network integration.To this end,this paper proposes a remote sensing image classification algorithm based on random multi-selective residual network.The algorithm integrates multiple residual networks to complete the classification task by multi-selection learning strategy.The algorithm sets an effective integrated learning objective function and uses the stochastic gradient descent algorithm to minimize the optimal classification error of each sub-network for each sample,and promotes the difference between each network.It can adapt to the classification task of a specific category,and form an effective classification.At the same time,its generalization is usually significantly better than a single learner.The effectiveness of the proposed algorithm is verified on two publicly available remote sensing datasets.Multiple residual networks canform optimal classifications for different types of remote sensing images,which can effectively improve the accuracy of classification.(2)A detection model for large-size remote sensing image is constructed in view of Yolov2 model,which can effectively detect ships and aircrafts objects.Object detection in remote sensing image faces some difficult problems such as complex background,diversified target scale and large image size.Aiming at these problems,this paper constructs a block detection algorithm based on Yolov2 model.By setting sliding windows of different scales,this algorithm can adapt to the scale changes of ship and aircraft targets in remote sensing images,and then establish the fusion of the block detection results in terms of non-maximum suppression.Thus,the large-scale satellite remote sensing images can be effectively processed.Experimental results verify the effectiveness of the algorithm.The effectiveness of the B-YOLOv2 algorithm is verified by comparison with the experimental results of Fsater R-CNN,SSD and YOLOv2 algorithms.