avinashingit/image-super-resolution

Project Implementation:

Method Used: k-nearest neighbors Hyperparameters: k: 1, patchsize: 4

Data Processing:

Loaded all the images one by one into the environment and converted them to numpy arrays. This is done for both low resolution and high resolution images. Each image array is now split into blocks of patchsize x patchsize. And then each block is mapped to the corresponding 2patchsize x 2patchsize block in the high resolution image array. Based on the location there will be 16000 blocks of each patch. A k-nearest neighbor classifier is fit to each of the each 16000 blocks. For example, if the patch size is 4, we would have 256 blocks and hence 256 arrays of 16000 blocks. For a given test image array, based on the location the train block at the same location which is nearest is found out. The corresponding high resolution block for the train block is chosen as the high resolution patch for the testblock. This is done for all blocks in a test image and for all test images.

Inspiration:

As x-ray images, unlike general images, don't tend to vary a lot, a lot of deep neural networks we tried didn't work out. This made us think simple and exploit the dataset properties and thus we came up with k-nearest neighbors classifier.