چكيده لاتين
Apple is one of the strategic products in Iran that has a significant share of agricultural exports. This product faces several challenges during the harvesting stages, such as physical damage and bruising, which is difficult to detect with traditional methods due to the lack of initial visual symptoms. These damages reduce the quality of the product and lead to mass spoilage of fruits in packaging, which results in significant economic losses. In this study, an attempt was made to identify the stains caused by bruising of apples in the early stages, which are not visible, by using near-infrared imaging and artificial intelligence algorithms, and to increase the shelf life of stored products by separating these apples. For this purpose, a database with 2000 images of local apples in the region was prepared.
Then, initially, in order to implement the algorithm on low-cost hardware with a limited database, a combination of traditional machine learning classifiers and manual feature extractors was used. These algorithms showed high accuracy of 99% and good generalizability. Then, deep learning algorithms were examined and the YOLO 8 Nano classifier model, which was able to identify damaged apples with 100% accuracy in the initial test, was selected as the base model. Considering the necessity of the algorithmʹs performance in different centers with diverse conditions, the generalizability of this model was evaluated and the results showed a decrease in the accuracy and recovery values to 84 and 74%, respectively. In order to improve generalizability, the base network was modified by adding a bidirectional feature fusion pyramid (BiFPN) structure and adjusting the learning parameters, which increased the accuracy and recovery by 5 and 15%, respectively. This modified model also performed better than the baseline model in detecting small, faint spots. Overall, this study showed that in addition to the possibility of using traditional machine learning models, adding feature fusion pyramid structures can enhance information retrieval and significantly increase the generalizability and sensitivity to small spot detection in the YOLO 8 classifier.
