An Entropy Driven Feature Selection Technique for Scene Text Classification using Crow Search Optimization

Abstract

Irrelevant and redundant features significantly degrade and impact classification performance in scene text recognition systems. Particularly, it become more impactful when handcrafted multimodal feature descriptor are employed for classification. Hence, this paper introduces framework including entropy driven Crow Search Optimization (CSO) for reducing high dimensionality of fused feature space and selecting salient features. The proposed technique formulates feature selection as a wrapper based naturally inspired optimization problem. The CSO in this work used to exploit social foraging behavior to iteratively refine feature subsets. A fitness function based on entropy computation is incorporated to quantify the feature information richness, discriminative behavior and relevance. The proposed framework is applied to serially fuse multimodal feature space generated from segmented Natural Scene Text (NST) images. These NST images are collected from challenging benchmark datasets SVT, MSRA-TD500 and KAIST. Extensive experiments demonstrate that the proposed framework significantly enhance the classification accuracy while reducing the computational overhead and false positives. The proposed framework is also compared against benchmark metaheuristic techniques including Particle Swarm Optimization (PSO), Genetic Algorithm (GA) and Ant Colony Optimization (ACO) in the same setup. Hence, the proposed framework confirms the stability, superiority and establish that entropy driven CSO as scalable and powerful feature selection strategy for complex scene text classification and recognition tasks.