Classifying Thai Occupation from Images using Deep Learning with Grayscale Feature Extractor

Abstract

Religious, ethnicity, gender, and occupation are some examples of social characteristics that can accurately define and explain human social behavior. Being able to determine people's jobs based on their visual information in photographs can assist with better identifying people, determining social roles, offering personalized recommendations, and conducting security investi-gations. In this paper, our goal is to extract occupational data from human clothing in images. We collected a dataset called TH-UniformDB, which comprises 10,000 photos of a single individual wearing Thai uniforms from nine occupation classes and the other class; each class has 1,000 images. The dataset exhibits a significant level of intra-class variety as well as inter-class similarities, which pose challenges in occupation classification. To address these issues and improve classification performance, we propose an approach that performs visual occupation recognition by combining the strength of processing the color images along with that of the grayscale features of the same images. According to our experimental results, the combination of grayscale and RGB features of images can effectively improve the recognition accuracy of the traditional deep neural network model between 3.15 to 10.15 percent, resulting in less impact of the inter-class similarity and intra-class variance.