In our work, we present a method by which we solve the problem of guided super-resolution of thermal images. The method is based on machine learning approaches using convolutional networks that are learned in a end-to-end manner. The model uses high-frequency RGB references to help reconstruct the high-frequency details of the output high-resolution thermal images. As source images, it receives realistic thermal images captured by a low-resolution thermal camera. The model has a simple five-layer structure with two separate entrances of convolutional layers. To capture the image collection, we assembled a sensor system of three cameras with which we captured images in the LWIR and RGB spectrum. Two collections of images were collected to evaluate the model, the first representing a limited light environment and the second representing an uncontrolled light environment. The collection had to be edited with the homography assessment method before learning. In the testing phase, we changed the sizes and number of filters in the convolution layers, as well as the learning data. Learning outcomes were evaluated with SSIM and PNSR image quality measures, which are widely used today. After learning the first collection, the model reconstructed images with an average SSIM value of 0.84 and an average PNSR value of 22.67 dB. When evaluating the reconstructions in the second database, the model achieved an average SSIM value of 0.62, and the average PNSR value reached 17,713 dB
|
