Sub-scene Target Detection and Recognition Using Deep Learning Convolution Neural Networks

Abstract

Sub-scene recognition algorithm based on super resolution along with scene dependent neural network model and sub-scene dependent target detection for automating object information extraction is proposed. This work deals with large number of challenges possessed by classification problems. Some of the challenges in this problem are the low resolution satellite images, diverse pattern of each sub-scene causing the low level learning for classification and plethora of distinct object classes present in each sub-scene causes low accuracy of object detection. Objective of this paper presents an image super resolution technique for rectifying problems posed by low resolution images with color density variations of chromaticity coordinates. To eliminate the problem of diverse patterns, have divided various land cover types into separate groups based on maximum mixed fraction among these groups and corresponding sub-scene recognition disparate model parameters are used to recognize various scenes. To increase the accuracy for object detection has developed a sub-scene dependent Neural Network model for extracting the target/anomaly of object information.