Automatic Detection of Flavonoids from Spectroscopic Images by Fusion of Two-Dimensional Convolution Product with Multi-Scale Textural Descriptors

In a context where traditional medicine occupies an important place in Côte d’Ivoire, the reliable identification of active compounds from medicinal plants is becoming essential. This thesis focused on flavonoids, known for their beneficial effects on health, and proposed an automated method for their identification based on the convolution product and texture descriptors. The main objective was to develop a simple, efficient, and reproducible method to automatically recognize these flavonoids from chromatograms obtained by thin-layer chromatography (TLC). For this purpose, two local plants, Paullinia pinnata and Morinda lucida, were studied. After extraction, the samples were visualized using specific solvents and reagents, then photographed under UV light at wavelengths of 366 nm and 254 nm. The images were processed in MATLAB using region segmentation and two-dimensional convolution algorithms. The results reveal a strong correlation between numerical parameters and flavonoid presence, with normalized Euclidean distances ≤ 0.05. This method eliminates subjective bias from visual analysis and validates the use of accessible tools for preliminary analyses. Limitations include the impact of image quality and residual spot overlap. This approach offers a fast, reproducible, and inexpensive alternative for flavonoid identification, suitable for resource-limited settings. Potential areas for improvement include the integration of AI (U-Net, Mask R-CNN) and the optimization of migration protocols.