Texture Image Classification Using Deep Neural Network and Binary Dragon Fly Optimization with a Novel Fitness Function

Abstract

Texture is one of the most significant characteristics of an image for retrieving visually similar patterns. So far, researchers utilize large number of gray scale features and several combinations of training and testing. As a result, they do not guarantee high accuracy due to mismatch between gray scale features and classifier type. With a view to develop a highly accurate system, the paper proposes a new approach using Binary Dragon Fly with Deep Neural Network based fitness function for texture classification (BDADNN). Contributions of the proposed BDADNN is twofold, first fusion of Fractal, GLCM and GLRLM features has been performed to incorporate border complexities as well as spatial dependencies of a texture image, then Binary Dragon fly Algorithm is applied for feature selection with a novel fitness function based on Deep Neural Network (DNN).The proposed fitness function is designed such a way that it captures maximum accuracy, relevance and mainly minimizing number of features and reduction among features. Deep Neural Network is utilized as a stack of auto encoders. Adapting five types of k-cross-validation (k $$=$$ 2,3,4,5,6 and 10) protocols, we classify the texture images using BDADNN. Our results demonstrate superior performance of proposed BDADNN compared to SVM, for all cross-validation protocols (K2, K3, K4, K5 and K10) in terms of sensitivity, specificity and accuracy.