An Efficient Super-resolution Algorithm for IR Thermal Images Based On Sparse Representation
W. Jino Hans, N. Venkateswaran · 2015
In this paper, an efficient learning strategy to super-resolve the down-sampled IR thermal image is presented.Though the resolution offered by state-of-the-art non-coolant based Focal Path Array (FPA) IR thermal imaging device is significantly high, the images captured by these devices has to be down-scaled for storing and transmitting it over a network.IR thermal images captured by non-coolant based FPA thermal imaging device is down-scaled by a scale factor and is represented as LR image.It is up-scaled by a simple interpolator such as bilinear interpolator to the desired magnification size to form the pseudo-HR image.Image patches are extracted from same location ( , ), from both LR and pseudo-HR image to form the self-example patch-pairs.Image patches which carry HF details are efficiently selected based on a threshold on sample maximum mean square error (SMMSE).Two effective dictionaries (LR and HR) constructed from patch-pairs are trained with state-of-the-art K-SVD algorithm.It is used within the sparse representation framework to reconstruct the up-scaled IR thermal Image.Quantitative and qualitative results claim that the proposed method super-resolve the existing LR images without introducing false HF details.