Many applications rely on thermal imagers to complement or replace visible light sensors in difficult imaging conditions. Recent advances in machine learning have opened the possibility of analyzing or enhancing images, yet these methods require large annotated databases. Training approaches that leverage data augmentation via simulated and synthetically-generated images could offer promising prospects. Here, we report on a method that uses generative adversarial nets (GANs) to synthesize images of a complementary contrast. Starting from a dual-modality dataset of co-registered visible and thermal images, we trained a GAN to generate synthetic thermal images from visible images and vice versa. Our results show that the procedure yields sharp synthesized images that might be used to augment dual-modality datasets or assist in visual interpretation, yet are also subject to the limitations imposed by contrast independence between thermal and visible images.

Thermal image formation can be modeled as the convolution of an ideal image with a point spread function (PSF) that characterizes the optical degradations. Although simple space-invariant models are sufficient to model diffraction-limited optical systems, they cannot capture local variations that arise because of nonuniform blur. Such degradations are common when the depth of field is limited or when the scene involves motion. Although space-variant deconvolution methods exist, they often require knowledge of the local PSF. In this paper, we adapt a local PSF estimation method (based on a learning approach and initially designed for visible light microscopy) to thermal images. The architecture of our model uses a ResNet-34 convolutional neural network (CNN) that we trained on a large thermal image data set (CVC-14) that we split in training, tuning, and evaluation subsets. We annotated the sets by synthetically blurring sharp patches in the images with PSFs whose parameters covered a range of values, thereby producing pairs of sharp and blurred images, which could be used for supervised training and ground truth evaluation. We observe that our method is efficient at recovering PSFs when their width is larger than the size of a pixel. The estimation accuracy depends on the careful selection of training images that contain a wide range of spatial frequencies. In conclusion, while local PSF parameter estimation via a trained CNN can be efficient and versatile, it requires selecting a large and varied training data set. Local deconvolution methods for thermal images could benefit from our proposed PSF estimation method.