Hyacinth Bean Quality Classification Using Pre-Trained Convolutional Neural Networks

The assessment of the quality of hyacinth bean provides grave problems in the agricultural production sector, where traditional methods have widely been used. These are time consuming and mainly subjective, and this augurs a high chance for human error. This problem thus creates an economic problem for farmers as it handicaps them to sale their produce effectively. There is a need for effective automated solutions on the quality categorization of hyacinth beans, which would ease the decision-making on better agricultural practice. The primary objective is to identify the most accurate and computationally efficient deep learning model for classifying hyacinth bean quality, ensuring scalability and real-world applicability. The research exemplifies the implementation of transfer learning in hyacinth bean quality classification using different deep architectures. The basic idea is the use and comparison of pre-trained models, to be specific, ResNet50, MobileNetV2, DenseNet121, NASNetMobile, EfficientNetB1, and EfficientNetB2 on a huge dataset of images of hyacinth beans. In this experiment, the architectures were fine-tuned and optimized to enhance quality classification performance. The dataset consists of a huge number of images representing different grades of hyacinth bean quality, thus forming a good basis for training and evaluation. The results showed that the top-1 accuracy of DenseNet121 reached its best value of 96%, whereas MobileNetV2 has demonstrated the highest computation efficiency under the value of 96%.