Volume Prediction of Axisymmetric Fruits Using Optimized Convolutional Neural Networks with Transfer Learning Strategy

Accurate volume measurement of fruits is key in the agricultural and food industries, supporting better logistics, quality assessment, and processing decisions. However, traditional methods for measuring volume are often manual and labor-intensive, creating bottlenecks in high-scale operations. This study presents a novel approach that utilizes convolutional neural networks (CNNs) combined with a transfer learning strategy to predict the volume of axisymmetric fruit from images, offering a more automated and efficient solution. To achieve this, five pretrained CNN architectures, including MobileNetV2, VGG-16, DenseNet201, ResNet50, and EfficientNetV2B0, were employed by modifying the fully connected layers and optimized through a random search process, allowing for optimal hyperparameter selection. Only the fully connected layers were fine-tuned, while the pretrained convolutional layers retained their original weights, enabling the models to focus on relevant image features without extensive retraining. The methodology encompassed dataset creation, image preprocessing, and segmentation, with training supported by the Adam optimizer and evaluated using mean squared error and mean absolute error. The performance of CNNs was assessed through metrics like mean absolute relative error (ARE) and the coefficient of determination (R²). Experimental results demonstrate that ResNet50 achieved the highest prediction accuracy with a mean ARE of 3.76% and an R² of 0.9721, outperforming other models and several existing methods from previous research. This study’s findings highlight the potential of CNN-based models, especially ResNet50, for precise axisymmetric fruit volume estimation. Future research may extend this method to encompass diverse fruit types and real-time applications, advancing automated processing technologies in the industry.