Integrating Shallow and Deep Features for Precision Evaluation of Corn Grain Quality: A Novel Fusion Approach

This study investigates the precision evaluation of corn grain quality, focusing on categorizing seeds into four classes: broken, discolored, pure, and silk cut. We evaluated 13 pre-trained CNN models, including AlexNet, VGG19, and ResNet, with AlexNet emerging as the top performer, achieving 71% accuracy validated through statistical analysis using Duncan’s multiple range test. To further enhance accuracy, we propose a novel fusion approach that integrates shallow features—extracted via 2D Discrete Fourier Transform (2D-DFT) and Hilbert transform—from the images with deep features from AlexNet. The deep features are combined with these shallow features to create an enriched feature vector. This vector, consisting of 1000 deep features, 140 2D-DFT features, and 140 Hilbert transform features, is classified using a Support Vector Machine (SVM). The hybrid model achieved an accuracy of 86%. Manual grain quality assessment is subjective and time-consuming; our automated framework addresses this challenge by providing a more objective and efficient evaluation. The integration of diverse features not only improves classification accuracy but also underscores the potential of combining various information sources for robust grain quality assessment.