PREDICTING VISUAL APPEAL IN ADVERTISING PHOTOGRAPHY

Abstract

Digital advertising has grown at a very high rate, supporting the necessity of images that are visually appealing and able to attract the consumer on the perception of the image and motivation to purchase a product. The study is a predictive model of visual appeal in advertising photography based on the methods of computational aesthetics, machine learning, and deep learning. The study conceptualizes, in the first instance, visual appeal as a construct of multidimensions that is influenced by composition, lighting, color harmony, subject prominence, emotional tone, and style. A large data set of advertising photos, gathered in several products and media are labeled by a structured labeling protocol, which measures aesthetic quality that is perceived by humans. They use both handcrafted and deep visual descriptors, obtained with the help of pretrained convolutional neural networks and transformer-based encoders, to construct predictive models. Its methodology consists of a preprocessing and normalization pipeline as a whole and two large families of models, CNNs to learn spatial features and transformers to learn contextual features worldwide. Empirical evidence shows that deep representations perform better than handcrafted features at fine-grain aesthetics and that transformer models are more able to predict the associations between the visual complexity and the scores of visual attractiveness. Another limitation noted in the study is the subjectivity of the datasets, cultural biasness, and lack of diversity in advertising situations.