<![CDATA[Fruit variety recognition using digital images is an important component in developing automated systems for agricultural handling and food-industry processing. The task is difficult because different fruit types may present nearly identical visual patterns, while image acquisition factors such as illumination, viewing position, and image quality can reduce classification reliability. To address this issue, this study designed a deep learning model for 131 fruit classes by adapting InceptionV3 through a transfer learning scheme. The pre-trained feature extraction layers were retained without retraining, while the original output structure was replaced with task-specific layers consisting of global average pooling, a 1024-unit dense layer with ReLU activation, and a softmax classifier. The image data were standardized to 224 × 224 pixels, augmented to increase visual variation, and divided into training, validation, and testing subsets using an 80:10:10 ratio. The proposed model produced an accuracy of 99.80%, with precision, recall, and F1-score values of 0.9900. These results exceeded the performance of GoogLeNet, ResNet, and VGGNet, showing that the use of pre-trained InceptionV3 features, customized classification layers, and augmentation can improve prediction consistency and reduce classification errors. Further evaluation on unconstrained real-world images and optimization for real-time use are recommended for future development.]]>
Copyrights © 2026
