Pattern recognition on aerospace images using deep neural networks

Pattern recognition is one of the most important tasks in aerospace image processing. Various methods based on convolutional neural networks attain state-of-the-art accuracy; however, their effectiveness on exact images is influenced by the chosen architecture and its training parameters.This work present methods based on convolutional neural networks for pattern recognition on the aerospace images. A possibility for objects segmentation into ten classes is demonstrated on example of the multispectral images from the World View 3 satellite. Four networks with different architectures were built, trained and optimized parametrically based on the auto-encoder neural networks. Segmentation results has been analyzed by means of three parameters: training Jacard Index, testing Jacard Index and weight numbers. The positive impact of the properly selected shearing augmentation on extension of a small marked dataset is discussed. The influence of the nonequilibrium classes on the segmentation accuracy and how to account this feature during training of deep neural networks is pointing out.