Automatic spice classification application by image recognition

Abstract

This thesis investigates the integration of shape recognition and deep learning techniques to develop a mobile-based spice classification system. Addressing the challenges of visual similarity and texture variability among spices, the research utilizes EfficientNetB3 as a transfer learning backbone, achieving a classification accuracy of 93.78%. The system combines Flutter for cross-platform frontend development with Django for backend processing, ensuring real-time performance and scalability. Key contributions include a curated dataset of 24 spice categories, optimized preprocessing pipelines, and mobile-specific model compression techniques. The thesis also discusses ethical considerations in AI deployment and potential future enhancements, such as augmented reality guidance. The results demonstrate that shape recognition can be effectively adapted for culinary applications, establishing a benchmark for mobile computer vision systems.