Hybrid-uniform mixture model based iterative robust coding for image recognition problems using thermal imaging

Abstract

In a human-robot coexisting environment, vision-sensing-based hand gesture recognition is considered a significant contemporary research problem in collaborative robotics. The recognition task becomes more challenging in complex environments, where accurate identification of specific hand gestures is crucial for correct robot navigation and task execution. The complexities in such challenging environments may include factors like improper illumination, presence of unwanted objects, occlusions, shadows, silhouette, and specularities. To address challenges in visual pattern recognition, regression analysis-based techniques have been widely utilized. These techniques aim to reconstruct a query error image or signal based on available training data. However, in the presence of substantial signal corruption, as mentioned earlier, traditional parametric and non-parametric statistical regression methods often yield inferior results. Robust regression approaches are specifically designed to mitigate the adverse effects of complex noises and challenges on the signal recovery process. These approaches enhance the system models to effectively attenuate the impact of such adversities. Furthermore, in this study, vein pattern recognition of the human forearm is also considered. Vein pattern recognition is an additional area of focus, where robust regression methods can be applied to improve the accuracy and reliability of recognition algorithms, particularly in challenging environments. In the context of the above discussion, this thesis investigates the application of a robust regression technique called Laplacian Uniform Mixture (LUM). LUM is a robust regression model based on sparse coding that is used for signal or image reconstruction. It is an enhanced version of the Regularized Robust Coding (RRC) model. The LUM model employs an iteratively reweighted regularized approach to achieve an optimal solution. The weight update process is guided by conventional logistic functions in the case of LUM. This thesis thoroughly examines the effectiveness of these techniques, particularly when dealing with training dictionaries that consist of real-world, corrupted hand gesture images. Environmental considerations, particularly challenging illumination circumstances that can compromise the system's functionality, can significantly impact the recognition ability of vision-based systems utilized in human-robot interaction (HRI). A solution based on far infrared thermal imaging is proposed in this study to mitigate this limitation of RGB cameras in collaborative robotics applications. The standard American manual alphabet (AMA) library from American Sign Language (ASL) is employed to interpret user input from humans and provide relevant guidance instructions. Sparse representation-based modeling is utilized to construct a robust framework for the entire system, incorporating advancements in image processing. A novel Student's t-uniform mixed error distribution model is introduced to accurately characterize the distribution of coding error residuals. By leveraging the properties of the Student's t-distribution, which exhibits a prominent peak at zero and elongated tails on both sides for higher coding residuals, the error distribution is effectively captured. Through extensive case studies involving normal-conditioned and occluded thermal sign images, the effectiveness of the proposed distribution model in far infrared vision-based sign language detection is demonstrated, even in the presence of environmental challenges such as low-light vision and block occlusion. Another work of this thesis explores the utilization of thermal images of the human forearm for vein pattern recognition. Two recognition methods, Laplacian Uniform Mixture (LUM) and the proposed Student's t Uniform Mixture (TUM), are investigated. The LUM method captures the complex distribution of vein patterns using a mixture model. However, limitations may arise in challenging lighting conditions or occlusions. In contrast, the proposed method, incorporating the unique properties of the Student's t-distribution, enhances robustness against environmental challenges. Extensive experiments on a dataset of thermal forearm images demonstrate the superior performance of the TUM method in terms of accuracy and robustness. This study contributes to the advancement of reliable and efficient biometric identification systems based on thermal vein patterns. Finally a retrospection of the entire thesis work is done, and the future research scopes are projected as well.