Customised ResNet architecture for subtle color classification

Abstract

This study addresses the challenge of recognizing subtle color differences, a problem critical to applications in fields such as healthcare, food production, and civil engineering. Specially research focusses on printed colors. The research evaluates multiple ResNet architectures, including ResNet-18, ResNet-34, and ResNet-50, to identify the most effective model for this task. Modifications to the ResNet-34 architecture are proposed, such as replacing average pooling with global max pooling and introducing max pooling layers within residual blocks, to enhance feature extraction and classification accuracy. The models were validated using a K-fold cross-validation, which confirms the effectiveness of the proposed approaches. The findings demonstrate the potential of these modifications to achieve high classification accuracy, showcasing their adaptability to real world scenarios. However, limitations such as the use of a specific dataset and the type of printer highlight the need for further research to generalize the approach across diverse datasets and conditions.