Executive Summary : | Seed quality estimation is a crucial area of research in the agricultural field, as it depends on parameters such as germination, physiochemical characteristics, defects, vigor, and disease. High-quality seeds are essential for plant growth and abundant harvest, and their use can lead to health risks. Seed adulteration affects the nutritional value and retail price of seeds. Traditional methods for seed inspection include protein electrophoresis, DNA molecular marker technology test, and high-performance liquid chromatography. However, these methods are destructive and time-consuming. Hyperspectral imaging is an emerging, non-destructive, fast technique that utilizes machine vision and spectral technology to acquire spatial and spectral information simultaneously. This technology has practical applications in agriculture, medical, and food sectors, including detecting seed quality, viability, defects, diseases, vigor, cleanness, moisture determination, and chemical composition. The proposed project uses two modalities for capturing images: RGB camera and Hyperspectral imaging camera. Spatial features will be captured from RGB images, while spectral features will be captured from hyperspectral images. A convolutional neural network for feature selection (CNN-FS) will be developed to select informative features from extracted features, with a vector of scores calculated to determine the importance of each feature on the final classification task. Another one-dimensional convolutional neural network with an attention mechanism (CNN-ATT) will be developed to classify seeds more attention to specific regions of input data obtained from CNN-FS. |