Design of computational intelligence based image restoration and classification algorithms

Abstract

ABSTRACT Image restoration and image classification are the two most fundamental aspects of image processing. Image restoration involves accurate preservation of colour, edge details, contrast, etc. of images which suffer from any sort of degradation either due to presence of noises like impulse noise, speckle noise, etc. or due to scattering and attenuation of scene light during inclement weather conditions. The scattering and attenuation of scene light increase during hazy weather condition due to the increase in the presence of dust, fog, mist and other aerosol particles in the atmosphere. Image restoration is necessary for various important applications like remote sensing, driver assistance systems, intelligent vehicles and for proper functioning of various computer vision algorithms. Image classification deals with automated recognition of class labels of different objects present in images. In order to design accurate as well as automated image classification methods, it is very essential to enable the designed methods to learn the most distinctive as well as significant characteristics of the objects present in images. These features are fed as inputs to the designed methods during the training phase in order to enable those methods to distinguish between different objects using training algorithms which are designed to artificially replicate the learning process of human cognitive system, so that these methods can henceforth perform classification of similar objects based on the knowledge acquired by them about the characteristics of the objects during the learning (training) phase. Image classification has found its use in medical field, remote sensing applications, etc. Image restoration deals with the preservation of lost details (color details/edge details/ contrast, etc. which are lost either due to noise corruption or due to atmospheric scattering and attenuation) in any degraded image and thus has to handle lot of uncertainties and randomness associated with the restoration process which arise due to the absence of true information about those lost details. Fuzzy Logic is one of the most important principles of Computational Intelligence which can effectively deal with such uncertainties and randomness. Image classification deals with the identification of the class labels of different objects present in images by learning their characteristics. Classification is usually done using neural networks which is also an important principle of Computational Intelligence. These networks are designed to artificially replicate the learning behaviour of human cognitive system. These networks differ from one another based on the number of hidden layers (the layers which lie in between the input layer and output layer) present in their architecture. Conventional neural networks usually comprise of lesser number of hidden layers while deep neural networks possess comparatively large number of hidden layers. The inherent feature extraction capability of deep neural networks have enabled them to improve their performance over conventional neural networks as it reduces the chances of erroneous classification which occurr due to use of handcrafted features in conventional neural networks. The above discussion clearly states the inter-relationship between Image processing and Computational Intelligence fields. In the present era, the principles of Computational Intelligence like Fuzzy Logic and Neural Networks have become the most favourable choices for the researchers to perform any image processing tasks as they can handle the fuzziness present in any data efficiently and can artificially replicate the learning behaviour of human cognitive system respectively. Novel methods for performing image restoration or image classification are designed in this thesis by exploiting the inter-connection between the Image processing and Computational Intelligence fields. Three real life problems are studied in this thesis: a. Daytime and Nighttime image dehazing b. Detection of structurally variant erythrocytes. c. Land cover classification using full-polarimetric image data These three problems have been chosen for study in this thesis because of their immense significance in real world. Detailed discussion on the practical significance of each of these topics is carried out in this thesis based on the statistical data available from various authorized sources. A comprehensive literature survey on each of these topics is presented in this thesis and the shortcomings of the existing works have been listed down and new methods are proposed to overcome the shortcomings. Some significant existing shortcomings of the chosen research areas which have served as the major inspiration behind choosing them are highlighted as follows: 1. Daytime and Nighttime image dehazing Although image dehazing is a well-defined research area and many researchers have proposed several methods to perform effective dehazing of images (particularly daytime images) but in most of those methods, the authors have performed image dehazing mostly protraying it as a simple contrast enhancement problem. They have failed to consider edge-preservation and noise detection which are other equally important and crucial aspects of image dehazing. Moreover, in most of the existing methods, image dehazing is performed assuming the nature of degradation to be similar across the entire image which in reality is not so as the nature of degradation varies region-wise across an image depending on many factors. So to perform effective image dehazing, there is a need of a method which performs dehazing considering the pixel-wise variation in the nature of degradation. In addition to these shortcomings, another significant limitation in this research area lies in the fact that the existing research in this problem area is mostly focused on dehazing daytime images. Although nighttime image dehazing is a very significant research area but very few works are done in that research area. No proper systematic survey is conducted in that field. No proper benchmark databases comprising of real world nighttime hazy images as well as their Ground Truth images exist in this field. 2. Detection of structurally variant erythrocytes The detection of structurally variant erythrocytes (poikilocytes) is an emerging research topic in medical image analysis field. Inspired by the real-life significance of this field many researchers have proposed various poikilocytes classification or detection methods but those methods mostly perform classification or detection either by counting the number of poikilocytes present in blood smear images or using handcrafted textural, geometric and shape based features. Methods designed in this field implementing the deep neural mechanism is very basic. Those methods simply use pre-trained networks to perform classification. This field currently lacks any deep neural architecture which can perform poikilocytes classification or detection using only significant and highly informative features from multiple networks. 3. Land cover classification using full-polarimetric image data Deep neural network based land cover classification is an upcoming research area in remote sensing field. The number of existing deep neural networks which are designed especially to perform land cover classification is very few and their performances is mostly tested on Synthetic Aperture Radar (SAR) images. As properties of SAR and Polarimertic Synthetic Aperture Radar (POLSAR) images vastly differ from each other, so designing a deep neural network especially for performing land cover classification using POLSAR images is required. Some notable contributions of the works proposed in this thesis are: - a. Bacterial Foraging and Fuzzy Logic synergism based daytime image dehazing method. b. Fuzzy Logic and Refined Colour Channel Prior synergism based nighttime dehazing method. c. An extensive survey work is conducted in this thesis focusing on nighttime image dehazing field. Some topics which are highlighted in the survey work are listed as follows: i. Dissimilarities between the characteristics of daytime and nighttime hazy images. ii. Unexplored aspects and real-life significance of nighttime image dehazing. iii. Inefficiency of the daytime image dehazing methods (methods which are designed based on the characteristics of daytime images) in performing nighttime image dehazing. iv. Existing nighttime imaging models and the differences between the models as well as their variations from the atmospheric scattering model (daytime hazy image model) v. Existing nighttime image dehazing methods and their limitations. This survey needs special mention as it is the first review work which is published in the nighttime image dehazing field. d. S-HAZE, a novel database which is created to benchmark the sky scene restoration capabilities of image dehazing methods. e. N-HAZE, a novel database. (This is the first database which is designed exclusively for benchmarking the performances of nighttime image dehazing methods.It comprises of hazy as well as Ground Truth images.) f. Daytime and Nighttime Dehazing Database (D&N-HAZE Database), a novel database. (This is the first database among all currently existing databases in image dehazing field that comprises of both daytime and nighttime hazy images of similar scenes, captured in the presence of real atmospheric haze as well as synthetic haze along with their corresponding Ground Truth images). g. An automated intensity based sky segmentation method. h. Novel feature ensemble creation method. (This feature ensemble method is designed to select the best features among Fully Connected (FC), Rectified Linear Unit (ReLU) and InverseReLU features to nullify the information loss occurring due to the suppression of negative values in features by the ReLU activation layer present in Convolutional Neural Networks. The efficiency of the proposed feature ensemble creation method is validated by performing detection of nine different types of erythrocytes having varied morphology using the features ensemble created by the proposed method.) i. Total Contribution Score Parameter (a novel feature selection parameter). j. Ranking method (a novel feature selection method). k. A novel Degree of purity & Scattering diversity based Advanced Lee filter. l. Novel deep neural network namely, Crop-Net for performing POLSAR image classification is designed. The utility of each of these proposed approaches in real world have been demonstrated and their efficiencies are validated by performing comparative analyses with the results obtained from existing methods. In most cases, the designed methods have achieved better performance efficiencies compared to existing methods.