| Popis: |
The recent advances in image capturing devices and their extensive use in real life scenarios have led to an increased availability of digital images on the internet. Efficient and effective image browsing and searching methods are required to retrieve and classify images in many fields such as computer vision, machine learning, pattern recognition, image processing, and signal processing etc. These fields specifically demand accurate and robust algorithms which can be used with vastly available computer systems and which are less time consuming. Over the past decade, significant attention has been paid to content-based image retrieval (CBIR) and image classification tasks. Extensive research has been conducted to develop local pattern-based techniques which address the problem of accurately classifying images based on the classes and CBIR tasks. However, these local pattern based techniques have many limitations when dealing with large texture datasets and they fail to achieve the task of adequately modelling image semantics, especially with regards to robustness against noise, response time with hardware limitations and most of all, retrieval accuracy. This thesis aims to present novel strategies for the analysis of texture datasets based on CBIR and classification of images, focussing on the investigation of vector quantization based strategies for the extraction of compact features and spatial information. Most of the techniques currently available are based on extraction of local patterns and usually the combination of these features leads to a huge size of feature vectors which affects not only the retrieval time but also retrieval accuracy. It is a very challenging task to develop such methodologies which address the issues of robustness and accuracy with compact feature size for fast processing. This motivation led to building accurate methods for publicly available texture datasets which can be used for object recognition and classification. Furthermore, the developed methods have shown massive potential when compared with deep learning architectures. In this thesis, two different content-based image retrieval methods to effectively retrieve texture images are proposed, as well as two image classification methods for face recognition, and an industrial project-based lobsters classification task. The first proposed method focuses on improving the local pattern-based techniques. Although these techniques are simple to use and implement for image retrieval tasks, they have less discriminative power and are more sensitive to noise due to their use of coarse quantization techniques. Moreover, description of a color image requires the combination of cross-channel information from all color channels of the image. The current practice of combining multiple channels is simply concatenating local patterns of channels, which not only loses cross-channel relational information but also increases the dimensionality of the generated feature vector. In order to overcome these disadvantages and capture more discriminative information among highly similar objects, a novel Multichannel Vector Quantized Local Pattern (MV QLP) method is proposed, in which a codebook is generated at each color channel to capture local correlated information among pixels with high discriminability to the textural variations among images and to preserve the information in quantized maps. For the combination of these quantized maps, a robust Multichannel Encoding Technique (MET) is proposed, to capture the cross-channel information with minimum feature dimensions. This method shows high discriminative power and robustness to variations in texture, color shades, and field of view of retrieving similar images from complex color image datasets. The next method for content-based image retrieval combines color vector quantization and visual primary features into a compact feature representation. Color vector quantization is proposed to describe the image in a compressed stream by preserving the contrast of an image, and to produce two color quantizers which are processed by vector quantization to preserve the content of a color image. The proposed extraction of visual primary features based on edge texture orientation and color moments features is loosely inspired by the human visual system and its mechanism for effectively recognizing objects in an image by its edges and color distribution. The representation of the proposed method utilizes histogram-based features which are populated by color quantization, and visual primary features that are used to measure the similarity between the twocolor images by a specific distance metric computation. The proposed method proves to be efficient and adaptive to the particulars of image retrieval, while not requiring any training information, making it suitable for real time color CBIR applications. Thirdly, a novel Local Mesh High-order Pattern Descriptor (LMHPD) for face recognition is proposed. This description is constructed in a high-order derivative space and is integrated with a Convolutional Neural Network (CNN) architecture. Based on the information collected at a local neighborhood of reference pixel with diverse radiuses and mesh angles, a vectorized feature representation of the reference pixel is generated to provide micro-patterns. They are then converted to multi-channels to use in conjunction with the CNN. The CNN adopted in the proposed architecture is generic and very compact with a small number of convolutional layers. However, LMHPD is derived in such a way that it can work with most of the available CNN architectures. To keep the computational cost and time complexity at the minimum, a lighter approach of high-order texture descriptor with CNN architecture that can effectively extract discriminative face features is proposed. Finally, the world's first method to automatically classify and investigate the moulting stage of Bay lobsters, formally known as Thenus orientalis, in a controlled environment is introduced. In this context, the image classification method only requires top view images of the exoskeleton of Bay lobsters. The texture of the exoskeleton was analyzed to categorize into normal, moulting stage, and freshly moulted classes. To meet the efficiency and robustness requirements of production platform, traditional methods such as Local Binary Pattern (LBP) and Local Derivative Pattern (LDP) with enhanced encoding scheme for underwater imagery were leveraged. Furthermore, a dataset was built of 315 Bay lobster images captured at the controlled under water environment for the classification task. |
