Machine-Learning-Based Image Feature Selection

Introduction

In the last few decades, the dimensionality of the data associated with machine learning and information mining errands has hugely expanded. Information with high degree and dimensionality has becomes major challenge with classical learning techniques (Belarbi, Mahmoudi, & Belalem, 2017). With the expansive number of features, a learning model can be non-fitted, and it results poor accuracy. To address the information mining challenges and issue of the FS, a wide range of research reference are available. Image feature descriptors selection is the best way to eliminate noisy and redundant set from feature to be used in the classification. Major purpose of any FS method is to choose a subset of features that reduce the redundancy and maximize the significance to the target set. FS is used for refining learning performance, reducing computation, and shrinking required space. To discriminate one class object from another for any pattern recognition algorithm features takes a major role. This chapter focuses on CBIR systems in which image features are extracted, and classified accordingly with one aspect of the relevancy of features with desired outcome. As a consequence FS plays a significant role CBIR and the improved selection process usually results in greater retrieval precision. Image classification is a generally contemplated issue in the analysis of images and also with, computer vision. Most of the classification frameworks can be separated into two major stages, include extraction of each image feature by a high-dimensional element vector. Next, these vectors are selected by help of classifier based on various search algorithms. Among feature extraction and the feature classification, an additional step can take place that is feature selection. This step, is about selection of a subset of features to improve the accuracy of information retrieval. Then again, include selection techniques have presently increased significant role to lessen semantic gap. Machine learning gives instruments by which vast amounts of information can be naturally examined. Feature Selection (FS) by distinguishing the most remarkable elements for learning, concentrates a learning computation on those parts of the information most helpful for investigation and future expectation. In such manner, this chapter is dedicated to show an idea of ways to deal with decrease the semantic gap using machine learning between low level visual components and irregular state semantics, through concurrent feature adjustment and highlight FS. To solve Feature Selection problem, a type of heuristic search algorithm can be used. Machine learning based search method is attractive intelligence optimization technique and many of the powerful method that has motivated and discussed in this chapter. Image feature subset selection is the way toward distinguishing and expelling however much immaterial and excess data as could reasonably be expected. This decreases the dimensionality of the information and may enable learning computations to work quicker and all the more adequately. Now and again, exactness on future characterization can be enhanced; in others, the outcome is a more reduced, effortlessly deciphered portrayal of the objective idea. On the basis of evaluation environments, FS procedures are divided into three major models as filter model, wrapper model and embedded model.