Deep learning models for brand analysis in visual imagery

Visual imagery published on social media provides a treasure trove of information that can be mined by brands and advertisement companies to understand how brands and products are portrayed. This thesis project aims at producing visualizations, based on established semiotic models, for product comparison. To achieve this goal, two components are required: adeep learning-based model for semantic understanding and a component for brand/logo detection. For the latter, brand logos and product names are extracted from advertisement images using deep learning techniques and the Google Cloud Vision API. To identify individual product names, even under noisy or cluttered image settings, the system combines logo identification with OCR-based text recognition, employing strategies such n-gram matching and Levenshtein distance. The goal is to guarantee very accurate brand and product recognition in actual advertising situations. Experimental results identified several detection challenges, such as small logo detection, multi-scale recognition, and few-shot logo recognition. Then, the semantic interpretation model aims at examining images’ implicit communication techniques by placing them inside a two-dimensional visualization map. The system learns to place each image by leveraging a multi-modal deep learning architecture that includes perceptron encoder. The proposed framework contributes to computational semiotics and advertising analysis by interpreting how brands create meaning in advertisements that goes beyond surface characteristics.