An effective image annotation using self-attention based stacked bidirectional capsule network

Abstract

This paper provides an advanced hybrid deep learning (DL) system for accurate image annotation. The first step consists of input images being pre-processed using three techniques: i) cross-guided bilateral filtering, ii) image resizing and iii) colour conversion to the green channel. After pre-processing, key features such as shape, wavelet and texture are extracted using three models: the modified Walsh-Hadamard transform, the extended discrete wavelet transform and the grayscale run length matrix (GLRLM). Once the feature is extracted, the optimal features are selected using the Chaotic Coati Optimization (CCO) algorithm due to feature dimensionality issues. After selecting optimal features, image annotation is performed using the self-awareness-based Stacked Bidirectional Capsule Network (SA_SBiCapNet) model. The stacking Bidirectional long short term memory model (BiLSTM) approach with a capsule network is applied to annotate the given images. The accuracy rate of the proposed method is 0.99 %. Therefore, the proposed method uses a hybrid DL model to perform effective image annotation.