Feature Pooling by Learning

Abstract

In learning-based image quality assessment, images are represented by features with low dimension much less than the size of image. The features can be obtained by the aid of priori knowledge that people have gained; for example, the aforementioned basic and advantage features. There is also increasing interest in learning-based features which are co-trained along with the learning tasks. For example, the so-called “deep learning” techniques are studied extensively recently to learn a task-oriented feature. Feature extraction and selection are performed to construct more efficient features of image by compressing the length of feature vectors in order to reduce computational complexity and, more importantly, to avoid overfitting risk as the small number of samples are used in training process. After feature extraction and selection, we need to map image features onto image quality value which is a real-scale number. This process is called “pooling” in the literature which is a kind of function of linear or nonlinear form. For example, summing up all quadratic components of a feature vector would come up with a real number that may represent image quality for some scenarios. This chapter encompasses several state-of-the-art pooling methods in machine learning approaches.