Comparison of Convolutional and Fully Connected Neural Networks for Durian Variety Classification

This study compares the performance of a Convolutional Neural Network (CNN) and a Fully Connected Neural Network (FCNN) in classifying seven durian varieties. A total of 1,825 images were used for training and validation in both models, which were evaluated based on accuracy, F1 score, and training efficiency. The CNN architecture, consisting of two convolutional layers, two max-pooling layers, and three fully connected layers, demonstrated superior performance. Its accuracy increased from 20.16% to 69.41%, while the average loss significantly decreased from 1.91 to 0.54. Additionally, the CNN completed training in 168 minutes and 11 seconds. In contrast, the three-layer FCNN showed an accuracy improvement from 30.33% to 54.58%, with a reduction in average loss from 1.94 to 1.29. However, its training process required 888 minutes to complete. These results indicate that the CNN not only achieved higher accuracy and faster convergence but also significantly reduced training time due to its ability to efficiently extract and process relevant image features. This study highlights the advantages of CNNs in handling complex visual patterns and emphasizes the importance of selecting appropriate neural network architectures based on data characteristics and performance objectives. The study demonstrates the effectiveness of deep learning models, particularly CNNs, in accurately classifying durian varieties, which can aid in automated agricultural processes and improve quality control in the durian industry.