Aperfeiçoamento do treinamento de redes de super-resolução deep learning a partir de imagens hiperespectrais aprimoradas

Abstract

Inserted in the context of Visual Computing and Remote Sensing, Super-resolution consists of the process of restoring high frequency information in low spatial resolution images. Traditionally, this type of technique seeks to solve the physical limitations of some imaging sensors in identify and analyze specific targets. With the increasing use of Deep Learning methods more robust approaches of Super-resolution has been gaining more and more space, as are the cases of Super-resolution networks based on Convolutional neural network. Although such an approach proved to be superior to traditional digital image processing techniques, mainly in RGB scenes, multispectral and essentially hyperspectral images need more attention by the method, since, by improving their spatial resolutions, the spectral consistency must be maintained, a fact considered one of the great challenges within the Super-resolution. Still in this context, due to the difficulty of obtaining hyperspectral images of low and high spatial resolutions properly registered, the scenes of low spatial resolution are synthesized from the process of degradation, resampling and noise of their corresponding of high spatial resolution . Although this flow is commonly adopted, it has not yet been evaluated what is the real influence that resampling techniques have on intelligent Super-resolution methods, since there is no consensus on the best technique to be used. Thus, the hypothesis is that the identification of the best resampling function of HIs de LR enables the Super-resolution models in Deep Learning to generate HIs of HR of better quality. Therefore, this work aims to evaluate resampling functions and identify the best function for the improvement of training of super-resolution networks in deep learning. As a proposal for this work, two different datasets of hyperpectral images consecrated in the literature were chosen and used in the process of synthesizing low spatial resolution images. Subsequently, with the data generated, their behaviors were evaluated within the best Super-resolution model selected based on the related works. This evaluation was performed from different metrics of comparison of hyperspectral images, especially the metrics: Peak Signal-to-Noise Ratio, Spectral Angle Mapper e Structure Similarity Index Measurement. From the values obtained hypothesis tests such as the case of the Friedman and Nemeyi tests were applied, in order to identify statistically which technique was best applied. Finally, the results obtained were compared and evaluated from a new set of data obtained in a controlled way, so that the spectral consistency could be evaluated based on predicted high resolution spectral images and point readings with non-iImageer spectrometer. From the results obtained, the resampling type Lanczos and Cubic presented the best results in relation to the others, thus proving, the hypothesis evaluated.