COLOR SPACE FUSION IN DEEP LEARNING MODELS FOR BIOMETRIC FACE IDENTIFICATION AND PLANT DISEASE CLASSIFICATION

Abstract

Visual classification is a challenging computer vision problem, with issues of lighting variation and color distribution, as well as feature representation. It has been applied in different sectors, such as biometric identification of faces and diagnosis of plant diseases.. Although effective feature learning is necessary for improving classification accuracy, some available approaches depend on unique color-space representations and fail to gather complementary visual data across domains. This is a problem since it is essential for improving classification accuracy. This study proposes a unified deep learning architecture that makes use of multi-color-space feature fusion in order to improve classification efficiency across a wide range of visual tasks. In order to construct a parallel feature-extraction framework that incorporates RGB, HSV, and YCbCr representations, the suggested approach makes use of ResNet-18 branches. To the best of our knowledge, this combination offers a succinct and domain-agnostic technique for multi-domain visual categorization. It does this by simultaneously accumulating luminance and chrominance characteristics, which strengthens the representation of discriminative features. For the purposes of biometric identification and illness classification, a dataset consisting of photographs of face features and plant leaves that was available to the public was employed. Following the completion of the typical preprocessing of the photographs, deep feature embeddings were retrieved from each color-space branch and then fused together through the usage of concatenation method. A number of performance criteria, such as accuracy, sensitivity, and specificity, were utilized in order to assess three different categorization situations. Single color space baselines, dual space combinations, and total fusion were the possibilities for these scenarios. The results showed that the classification was better when multi-color representation (MC) was used. The lowest accuracy was obtained for face recognition with 52.43% accuracy while the maximum accuracy was obtained for plant disease classification with 94.50% accuracy. For a number of statistical studies, scenarios, it is possible to have different color representations and categorize the images correctly. The proposed approach is scalable, generalized and is computationally efficient. This approach may be used to enhance CNN classification systems in all kinds of imaging applications