AI-Based Image Classification System Using Transfer Learning and Interactive GUI

Abstract

The rapid advancement of Artificial Intelligence (AI) and Deep Learning has
significantly enhanced the capability of machines to interpret and understand visual data.
Image classification, a core task in computer vision, plays a vital role in numerous
applications such as medical diagnosis, surveillance, agriculture, and autonomous
systems. This project presents an AI-based Image Classification System that leverages
transfer learning techniques combined with an interactive graphical user interface (GUI)
to provide an efficient and user-friendly solution. The proposed system utilizes a pretrained
deep learning model, integrated through a custom classifier module, to classify
input images into predefined categories. Transfer learning is employed to reuse
knowledge from large-scale datasets, reducing training time and improving accuracy
even with limited data. The system is implemented using Python, with libraries such as
TensorFlow, NumPy, and PIL for image processing, while CustomTkinter is used to
design a modern, responsive GUI.
A key feature of this system is its asynchronous processing capability. The model is
loaded in a separate thread to ensure that the user interface remains responsive, thereby
improving user experience. Users can upload images through a file dialog, view them
within the application, and initiate classification with a single click. The system then
processes the image and displays the top predictions along with confidence scores. To
ensure compatibility across different environments, especially between modern NumPy
versions and older TensorFlow implementations, necessary patches are applied. This
enhances the robustness and portability of the application. The GUI is designed with a
sidebar for controls and a main content area for image display and result visualization,
ensuring intuitive navigation. The results demonstrate that the system can effectively
classify images with high accuracy and minimal latency. The modular architecture allows
for easy extension, such as integrating new models or expanding the dataset. This makes
the system adaptable for various real-world applications.