|M.Sc Student||Shadhan Lior|
|Subject||Detection of Anomalies in Texture Images Using Multi-|
|Department||Department of Electrical and Computer Engineering||Supervisor||PROF. Israel Cohen|
|Full Thesis text|
Anomaly target detection is the process of locating elements in a scene which are unlikely to be a part of it. This is a challenging problem, mainly due to the large variability of the scene’s background clutter and the appearance of the anomalous elements. The detection process is generally performed using a Bayesian classifier with respect to a predefined probabilistic model and an appropriate feature space in which a clear segregation between the anomalous elements and the rest of the background clutter in the scene is possible. Texture segmentation is the process of segmenting different textures in a scene based on a texture classification scheme. Texture classification is generally performed with respect to a characterizing feature space and a distance measure between textures. Since images of the same underlying texture can vary significantly, textural features must be invariant to image variations and at the same time sensitive to intrinsic spatial structures that define textures. Furthermore, the distance measure must be robust to these variations in order to avoid classification errors. As a result, texture segmentation algorithms perform poorly when used for detecting anomalous targets in a given scene.
In this work, we integrate texture classification features with properly formulated anomaly detection classifiers, achieving improved performance in both realms. We introduce a multi-resolution feature space which follows the Gaussian distribution and corresponding algorithms for anomaly detection and texture classification. The proposed feature space is based on a multi-resolution representation of the image, obtained using the redundant discrete wavelet transform (RDWT), followed by estimation of Gaussian scale mixture hidden multipliers and certain non-linearities which improve texture classification performance. Detection and classification are then performed using a single hypothesis test classifier. In addition, we introduce a multi-resolution random field model (RFM) and a corresponding anomaly subspace detection algorithm designated to detect additive anomalies in harsh environments where signal to noise energy ratio values are very small. The proposed model is based on a multi-resolution representation of the image, obtained using the RDWT, followed by a squaring non-linearity. Each layer is modeled as an RFM with different sets of parameters. Detection is then performed using a matched subspace detector classifier, formulated for detecting subspace signals in RFM innovations of the background clutter. The proposed algorithms are shown to have improved performance compared to those reported recently in the literature. The results demonstrate the robustness and flexibility of the algorithms in adverse environments.