ACCURATE IMAGE CLASSIFICATION FOR IMPROVED VISION RECOGNITION USING HIERARCHICAL TRANSFER LEARNING

Abstract

Image classification is an important task by computer vision with applications like object recognition to medical diagnosis. In last few years, transfer learning has emerged as effective technique to improve the image classification performance by utilizing pretrained models knowledge on large scale. The research proposes a novel method for image classification using transfer learning to harness the hierarchical relations classes and enhance the accuracy of vision recognition system. A most suitable dataset is collected and then preprocessed, including steps such as normalization, resizing, data cleaning and argumentation. The design and implementation is done using hierarchical transfer learning framework. The trained model is evaluated on performance metrics like precision, accuracy, loss compared against baseline model to measure the improvement achieved. The Python programming language, along with the TensorFlow framework, and Google Colaboratory hardware, was utilized for this thesis. Existing models available online were selected and modified accordingly. The proposed methodology harnesses the hierarchical structure, resulting in enhanced accuracy for image classification tasks. The research approach presented in this study contributes to the progress of vision recognition systems and holds potential applications across diverse domains, such as medical image analysis, object recognition, and autonomous systems.