Improved sentiment classification of images using a two-stage relieff and modified SCA feature selection

Visual media possesses rich semantics and a remarkable ability to convey emotions and sentiments. This work, aims to recognize the sentiment conveyed by the visual content of an image. This task is quite complex as it requires extracting high-level abstract content from visual data. Both global and local regions within an image carry significant emotional cues. A saliency-based approach is utilized to predict visual attention, highlighting the most crucial local areas in an image. The proposed framework consists of two main modules: (1) feature extraction from global and local image regions using a pre-trained Darknet53 CNN, which captures high-level concepts, and (2) a two-stage feature selection process using a modified Sine Cosine Algorithm (SCA) and the ReliefF algorithm, followed by classification with an SVM classifier. The proposed Modified Sine Cosine Algorithm (MSCA) enhances the search path of the original SCA by introducing a new convergence empirical parameter, preventing the algorithm from getting trapped in local optima. The experimental findings, evaluated on four datasets, demonstrate that the bi-stage feature selection combined with a deep learning approach significantly improves the accuracy of sentiment classification.