Abstract

Due to the growing increase of generated images via cameras and various instruments, image processing has found an important role in most of practical usages including medical, security and driving. However, most of the available models has no considerable performance and in some usages the amount of error is very effective. The main cause of this failure in most of available models is the distribution mismatch across the source and target domains. In fact, the made model has no generalization to test data with different properties and distribution compared to the source data, and its performance degrades dramatically to face with new data. In this thesis, we propose three novel approach which is robust against data drift across domains. Our sparse representation and subspace alignment (SRSA) approach is a novel image optimization problem based on the combination of instance-based and feature-based techniques. Our proposed approach selects source samples that are relevant to target samples using sparse representation in first phase. Then, SRSA maps the source and target data into their respective and independent subspaces. In last phase, SRSA aligns the mapped subspaces to reduce distribution mismatch across domains. The proposed model Sparse coding and ADAptive classification (SADA), reduces the distribution difference across domains via generating a common subspace between the source and target domains and increases the performance of model. Also, SADA reduces the distribution mismatch across domains via the selection of the source samples which are related to target samples. Moreover, SADA adapts the model parameters to build an adaptive model to encounter with data drift. At the end, we propose a method for Prediction Error Minimization of Image Classification Models via Sparse Coding and Domain Adaptation (SCDA) that exploits sample reweighting to select source domain samples that are related to target data. Furthermore, SCDA preserves the intrinsic structure of data in new subspaces, and also reduces the conditional distribution difference and aligns projected subspaces. Our variety of experiments demonstrate that the proposed approaches outperforms state-of-the-art domain adaptation methods.

Key Words: Image processing, Sparse coding, Visual domains adaptation, Adaptive classification.