Hybrid Deep Learning–Machine Learning for Bird’s Eye Chili Quality Classification

Abstract

The manual inspection of dried bird’s eye chili (Capsicum frutescens L.) is prevalent yet susceptible to subjectivity, inter-rater variability, and low efficiency. This research introduces a hybrid deep learning and machine learning pipeline for the classification of images into three categories: fresh, medium, and dried. The method encompasses staged image acquisition throughout the drying process, preprocessing (including HEIC to PNG conversion, background elimination, scaling, and normalizing), and real-time augmentation to enhance robustness. Feature embeddings are obtained from MobileNetV2 by transfer learning utilizing a Global Average Pooling head and are evaluated against EfficientNetB0, NASNetMobile, ResNet50, and DenseNet121. The embeddings are categorized using various algorithms: Random Forest (RF), Support Vector Machine, and a shallow Artificial Neural Network, with RF selected for its consistent performance. Evaluation employs an 80/20 split, focusing on accuracy, precision, recall, F1-score, and confusion matrix analysis. Results indicate that MobileNetV2 produces the most distinctive features, whereas RF provides the most reliable downstream predictions: the system achieves 94% validation accuracy during feature extraction and 91% at the final classification stage, alongside high precision-recall and minimal misclassification. The chosen MobileNetV2+RF model is implemented in an Android application for real-time inference from smartphone photos, providing class labels and a moisture-level signal based on mass-loss measurements to aid postharvest decisions. The contributions consist of an objective and efficient quality-assessment pipeline, an empirical model comparison, and a deployable mobile implementation. Future endeavors will focus on extensive datasets, multimodal signals, and cross-variety generalization.