Created deep learning based gradient image super-resolution network with faster convergence and memory optimization to preserve SIFT features for recognition of low-resolution images; Super-resolved Difference of Gaussian (DoG) images are integrated to SIFT which exhibits around 10-12 SIFT matching points gain over the state of the art super-resolution method EDSR.
Using Deep Learning, middle frame from a video is predicted from the previous and following frames and the result exhibits a 3dB gain over bi-cubic interpolation.
DNA sequence being mapped into codon level and then converted into character text data are modelled and mapped to form a text file for the learning based compression, e.g., PAQ which is the combination of neural network text prediction and arithmetic coding.
Using GPAC tools, same video with different resolution is converted to different bit rates. For each one, the video is divided into different segments,each denoting a video content. From each segment, image frames are extracted using FFMPEG. For memory optimization, only the top features with larger eigenvalues from SIFT are taken applying a dimensional reduction by PCA and then the Fisher Aggregation is done for different GMM models.