Deep feature based symmetric positive definite representations for image recognition

Almost in every computer vision algorithm, the most important task is to extract the relevant information. The main challenge in this task is that the images cannot be captured in a controlled environment. There will always be problems such as illumination variations, occlusion, background cluttering, viewpoint differences. Therefore, computer vision applications require robust and discriminative image representations in order to extract the relevant information. Designing such robust image representations has always been a challenge. For many years, the state-of the- art computer vision algorithms were based on so-called shallow or hand-crafted image representations which are directly computed from the raw pixel data. Even though these features are computationally cheap, they require extensive knowledge about the data in hand and the discriminatory power is unsatisfactory.