Multi-Scale CNN Architectures for Enhanced Feature Extraction in Image Sentiment Analysis
S Indhumathi, F. Mary Harin Fernandez · 2025
Imagery sentiment analysis is a complex task, the aim of which is the interpretation of both local and global visual signs in order to correctly classify the emotions ordered by the image. CNNs old architectures commonly have trouble with the proper balance of the tip-oriented extraction and the overall understanding of the context. This research suggests a new multi-scale CNN architecture in order to improve the extraction of features by the hierarchical representations of the visual sentiment cues. The architecture operates by adding convolutional layers of nucleotide scale and pyramid pooling modules to take features from different spatial resolutions. Besides, the attention mechanism is applied to the sentimental areas, thus dealing with the noise from unimportant parts. That is modeling a CNN with multi-scale convolutional layers in the construction of an image where each kernel size will deliver what is a fine signal versus a contextual pattern. A pooling module based on a pyramid structure is collected to the resolution to obtain features from different spatial scales, so the model will have the view of the entire image at a context level. Then, in addition, the proposed attention mechanism that finds and highlights regions with more probability of emotional content is used. The suggested architecture is prepared and tested on a number of datasets, that reflect the real-world sentiment contexts and the evaluation metrics like accuracy, precision, recall, and F1-score, which show the effectiveness of the proposed architecture. The proposed approach seeks to give a more in-depth definition of visual sentiment which, in turn, will lead to more precise classification as well as allow for better understanding with attention-based visualizations. The results are anticipated to drive the image sentiment analysis field further, making it relevant for social media analytics, advertising, and human-computer interaction.