Design of a high-density bio-inspired feature analysis deep learning model for sub-classification of natural & synthetic imagery

Differentiation between Natural and Synthetic imagery is a specialized area of digital image processing that assists in identification of computer-generated entities. But due to advancements in computer graphic interfaces, it is difficult to identify synthetically generated images from their natural counterparts. A large number of deep learning models are proposed by researchers, that perform complex feature analysis in order to differentiate between these image sets. But these models are either highly complex, or do not process sub-components of the image, which results in reduced efficiency under large-scale scenarios. Moreover, these models showcase limited accuracy when used in categorical classification modes. To overcome these issues, this text proposes design of a novel high-density bio inspired feature analysis deep learning model for sub-classification of natural & synthetic image sets. The proposed model initially uses a YoLo model to identify different objects in the input image sets. These objects are individually processed via a hybrid LSTM (Long Short-Term Memory), and GRU (Gated Recurrent Unit) model, which assists in estimation of high-density feature sets. These feature sets are further extended using Wavelet, Cosine, Fourier & Convolutional features for multi-domain representation of input images. These feature sets are given to an Elephant Herding Optimization (EHO) Model, which assists in identification of feature sets with high inter-class variance levels. The selected features are classified via a customized 1D CNN model, and validated via an efficient GAN classifier, that assists in classification of these sub-entities into natural & synthetic objects. These classification results are combined in order to identify if the input image is purely natural, purely synthetic, of a mixture of both types. The model also evaluates a Natural Image Score (NIS), which assists in estimation of natural content in each of the input image sets. It was observed that the proposed model showcased 8.3% higher accuracy, 10.5% higher precision, 3.9% higher recall, and 8.5% higher AUC (Area Under the Curve) performance when compared with standard classification models under real-time scenarios.