Cognitive assisted deblur enhancement for images restoration
Publication Date: 2017-Jun-01
The IP.com Prior Art Database
TITLE: Cognitive assisted deblur enhancement for images restoration Abstract
Disclosed is a method to apply to blurred images as counter-measure to blurring, in order to improve the image sharpness combining state-of-the-art deblur techniques and state-of-the-art cognitive visual recognizers. First a visual recognizer is used to classify a large training set of known blurred images that can be fully recovered then an image recovery tool is trained to recover specific scenes. A trained anti-blur recovery tool that combines the best recovery tool for the recognized scene can be adopted at runtime by final users.
Using hand-held cameras, it is possible that resulting images are affected by blurring caused by multiple reasons, including inadequate light for the scene hence driving long exposure times comparing to the relative movements of the objects being pictured. There are several existing image deblur techniques available, ranging from hardware or software systems, with advantages and drawbacks of individual techniques. Most of these techniques are useful during the taking of the image, other can work statically on existing images using image processing algorithms but with limited successes. This article describe a method to enhance (via state-of-the-art cognitive visual recognition systems) the effectiveness of a set of old and new techniques to solve this problem: starting from a blurred picture, obtaining an image as close as possible to the original image a person could have taken if he was able to get a picture of a steady object or person or landscape. The core idea is to leverage visual recognition techniques to obtain detailed image classifications and image area separation (e.g. human subject, car, sky, clouds, moon, buildings, etc.). Starting from a training data set of unblurred (perfect) pictures (representing various subjects), state-of-the-art visual recognition tool run to get area classifications of each picture, and add metadata to the same picture to list these extracted cognitive information. Starting from the same set of perfect pictures, synthetic data set is generated simulating a motion issue in each picture. This data set can be generate defining a simulated – but realistic – motion trajectory of camera (or smartphone) and applying it of original data set, generating a huge set of blurred images, each one containing a different combination of simulated motion trajectory and velocity gradients. Normally only the minimum blurring caused by small limited movements/oscillations, typical in smartphones, is to consider to get the maximum effect in restoration. Then this new generated data set will have the huge advantage that, keeping the simulated movements small enough, we can clone the metadata information computed from the each original photo picture.
At this point we have a new huge data set of metadata-enriched blurred image and it is possible to train again state-of-the-art visual recognition tool assigning the same t...