Deep feature extraction based WaveNet CNN framework for visual sentiment analysis

The rise of internet reviews, driven by smartphone use, has led to a trend of incorporating images in feedback on platforms like Yelp, TripAdvisor, and Google. This research shifts focus from traditional textual sentiment analysis (SA) to visual sentiment analysis (VSA), particularly in the restaurant domain, where images of food, ambiance, and facial expressions significantly influence user perception. People often check reviews before selecting a restaurant for dining. Reviewers frequently post restaurant images with the purpose of reviewing them in order to classify the polarity of sentiment depending on factors such as face expressions, food presentation, or ambiance. Therefore, this research focuses on analyzing images to classify them in neutral, negative, or positive sentiment categories. This study addresses limitations in existing VSA approaches, particularly in extracting and selecting discriminative features for accurate sentiment classification. The framework processes images through a combination of a custom convolutional neural network (CNN) and an optimal wavelet transform, utilising spatial and frequency features. To optimise feature selection, a dual-moth flame optimization (DMFO) algorithm is employed, followed by an ensemble classifier for sentiment prediction. The proposed approach achieves an impressive accuracy of 75.15%, 87.42% and 89.72% on the FER2013, JAFFE and restaurant review datasets, respectively, outperforming existing methods. The results demonstrate the effectiveness of combining multi-domain features with optimized selection techniques. This framework can support real-world applications such as restaurant decision-making, customer experience analysis, and automated review understanding.