Hybrid Deep Learning with Image Segmentation Guided Visual Analysis for Multiclass Millet Grain Classification in Smart Agriculture

Abstract

Classification of grain via traditional image-based methods for agriculture is truly challenging. Variations in texture of the grain, lighting conditions, or perhaps background noise would often prevent an image classifier from learning and generalizing well. This becomes particularly evident with grains that present a highly similar feature distribution. To solve such problems related to grain classification, a novel hybrid deep learning-based framework using image segmentation techniques has been proposed to enable the finer and precise classification of millet grains. Systematic preprocessing in the form of resizing, normalization, and dynamic augmentation was inculcated from the publicly available millet dataset with the intention of imparting more robustness to the models. Several CNN architectures such as MobileNetV2, VGG16, DenseNet169, Xception, and their respective hybrids were trained. Segmentation techniques such as Otsu Thresholding, KMeans Clustering, Watershed, and SLIC Superpixels were implemented. Among the different models, the hybrid model DenseNet169 + Xception achieved the highest results, attaining a higher than 95% accuracy along with class-wise precision, recall, and F1-score. These findings show that a merger of segmentation with hybrid deep learning can yield a more explainable and accurate classification of grains, providing a strong base for smarter agricultural solutions.