Implementation of the DenseNet201 Architecture with Transfer Learning for Web-Based Shallot Quality Classification

Shallot quality is one of the important factors affecting product value and market acceptance. However, the quality assessment process is generally performed manually through visual inspection, making it prone to subjectivity and inconsistency. This study aims to design and implement a web-based shallot quality classification system using the DenseNet201 architecture with the Transfer Learning method and to evaluate the effect of batch size variations on model performance. The dataset consisted of 1,200 shallot images divided into two classes: good quality and poor quality. The research stages included data collection, image preprocessing using Contrast Limited Adaptive Histogram Equalization (CLAHE) and sharpening, dataset splitting, model training using DenseNet201 Transfer Learning, and system implementation using Laravel and Flask frameworks. Experiments were conducted using three batch size configurations: 16, 32, and 64. The training results showed that all configurations achieved excellent performance, with validation accuracy reaching 99.44% at the end of the fine-tuning stage. During testing, the model with a batch size of 32 achieved the best performance, obtaining 100% accuracy, 100% precision, 100% recall, 100% F1-score, and a loss value of 0.0456. Furthermore, the developed system was able to classify shallot images automatically through a web interface with an average inference time of 294 ms per image. The results demonstrate that the combination of DenseNet201 and Transfer Learning is an effective approach for shallot quality classification and has the potential to support faster, more consistent, and objective quality assessment processes