Canned Apple Fruit Freshness Detection Using Hybrid Convolutional Neural Network and Transfer Learning

Abstract

Fruits contain good nutrients such as protein and vitamins; therefore, fruits are usually used as complementary food ingredients. Apple fruit is one of the most important traditional table fruits in the temperate zone besides being the most commonly consumed fruit in the world. Apple’s freshness is a decisive characteristic of consumer choice. Consumption may increase if apples of satisfactory freshness and quality are provided to the consumers. Reduced freshness of canned apple fruit results in significant spoilage during storage and distribution, affecting both agricultural profitability and consumer satisfaction. Many approaches have been presented to predict apple fruit freshness; however, most of these studies suffer from low accuracy due to poor image quality and the unavailability of sufficient data. This study aims to grade the freshness level of canned apple fruits by combining digital image processing and machine learning techniques. The proposed technique involves data collection of apple fruit, image segmentation, and preprocessing, creating the neural network model, training the network model, and testing for apple freshness prediction. A dataset is collected for apple fruit images for fresh, semifresh, and rotten classes, which is further augmented for more images. We extensively evaluated the performance of the proposed convolutional neural network–transfer learning (CNN-TL)–based technique. For performance evaluation, ResNet, GoogleNet, AlexNet, VGG, MobileNetV2, and InceptionV3 are also utilized due to their superior performance reported in the existing literature. The experimental results indicate the superior performance of the proposed approach with a 98% accuracy on the original dataset, which is better than other deep learning models. The model secures a 97% accuracy on the augmented dataset. Cross-validation results report a 97.71% accuracy using five folds. Furthermore, a comparison with existing state-of-the-art approaches shows that the proposed approach has better results for apple freshness classification.