INTELLIGENT IMAGE PROCESSING WITH DEEP LEARNING

Abstract

Artificial neural networks have been recognized for a number of years as a powerful tool for solving real world image processing applications. These networks are able to deduce meaningful interpretation from easily available sample data. They can extract patterns and learn trends that are too complex to be captured by conventional analytical approaches. Interest in neural networks research has historically seen many ups and downs. There has been remarkable improvements in deep learning in the last decade. It has now been proved that deeper networks can approximate much deeper unknown functions. They have been vastly applied in speech and image recognition problems. In this dissertation we focus on applying deep neural networks to some real world image processing applications.

One of the areas we target is the compression of digital images which refers to the encoding of images with as few bits as possible. A few studies have been carried out for adapting the artificial neural networks for digital image compression. One of the issues with these network is the time required for their training. We attempt to reduce the training time of neural network by employing neurons which are more biologically plausible and their activations are very efficient to compute. By employing these neurons namely the rectified linear units (ReLU's) the training time for the deep neural network is significantly shortened. The network also shows better generalization property and the PSNR of the reconstructed images is proved to be better than the conventional logistic sigmoid neurons.

The other application we target in this dissertation by the deep neural networks is the visibility enhancement of scenes in an image which are degraded by fog. Many solutions have been proposed by various researches for the recovery of image scenery from images acquired during bad weather conditions. Mostly these solution rely on complex mathematical models with a lot of parameters and heavy filtering techniques to subtract the weather effects and restore the original scenery in an image. We propose a solution to recover the scene from a foggy image by learning the corresponding fog function by a deep neural network. The deep neural network learns this fog function from various foggy images and then removes it from the images for better visibility. The experimental results proved to be effective for a large set of unseen images.

The deep neural networks are meant to explore and learn some underlying relationships/trends/pattern present among the data. It is possible to employ a previously learnt feature to help master a new task a technique known as transfer learning. We explore this concept to design an efficient encoder (i.e. a deep neural network) for compressing totally unrelated images. We take advantage of the fact that the these different scene images have correlations at very fine scales. This correlation allows us to share the learning experience gathered by one task to another task. By applying this transfer of parameters we were able to compress unrelated images efficiently.

Furthermore deep neural networks are algorithms which exhibit a lot of parallelism. To take the advantage of this parallelism and further fasten the speed of these deep neural networks we implemented the proposed architectures on a graphics processing unit (GPU). This implementation yielded in a speed up of several folds for our proposed deep neural networks.