Multi-Class Millet Classification Using A Fusion Deep Learning Convolutional Neural Network

Abstract

Reliable identification of millet cultivars is essential for maintaining grain quality, supporting seed authentication, and improving automation in post-harvest processing. Despite recent advances in computer vision, accurate classification of millet varieties remains challenging because many cultivars exhibit subtle visual differences in color, texture, shape, and grain arrangement, particularly when analyzed in bulk grain images. To address this challenge, a fusion-based convolutional neural network is developed that combines feature representations extracted at multiple spatial scales. The proposed architecture employs parallel convolutional branches with different receptive fields to capture local texture patterns, structural grain characteristics, and broader spatial relationships within bulk grain formations. The extracted features are integrated through a fusion mechanism to generate a richer and more discriminative representation for classification. To enhance model robustness under practical imaging conditions, image augmentation techniques including contrast variation, noise injection, sharpening, and rotation were incorporated during training. The proposed model was evaluated against AlexNet, VGG16 trained from scratch, pretrained VGG16, and pretrained InceptionV3 using the same experimental settings. Among all evaluated architectures, the fusion network consistently achieved the best performance, attaining a validation accuracy of 93.80% and a testing accuracy of 93.09%. The model further achieved a Precision of 93.02%, Recall of 92.04%, and F1-score of 92.52%, indicating balanced and reliable classification performance across millet cultivars. Analysis of the confusion matrix revealed strong class-wise discrimination with only limited confusion among visually similar varieties, while cultivars possessing distinctive color characteristics were identified with perfect accuracy. The results highlight the importance of multi-scale feature learning for fine-grained agricultural image classification and demonstrate that feature fusion can substantially improve cultivar recognition in bulk grain imagery. The proposed framework offers an effective solution for automated millet identification and can support the development of intelligent grain sorting and quality assessment systems for modern agricultural applications.