Specular Correction Of Images Using Autoencoder Neural Network Model

Abstract: A method (600) and a system (100) for correcting set of input images is disclosed. The computing device (102) receives set of input images from an imaging device (112). A corresponding set of corrected output images are generated based on processing of each of set of input images using pretrained and unsupervised autoencoder neural network model (300). For each of set of input images, the corresponding input image is encoded to generate encoded image data by extracting image parameters by set of encoding layers (304). The encoded image data is decoded to generate corresponding output image by set of decoding layers (308). Further, one or more regions of interest corresponding to specular reflection in the corresponding input image are determined by determining a noise mask. The one or more regions of interest are reconstructed in corresponding input image by predicting one or more pixels of region of interest. [To be published with FIG. 3]

Patent Information

Applicants

Inventors

Specification

Documents

Application Documents