Lemon Images Classification Using Efficientnet and Resnet-50 Models
No Thumbnail Available
Date
2023
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Ülkemiz limon üretiminde dünyanın önde gelen ülkeleri arasındadır. Araştırmalara göre dünyadaki limon ihracatının dörtte biri Türkiye tarafından yapılmaktadır. Bu kadar yüksek üretim kapasitesine sahip olan ülkemizde limon sınıflandırması birçok alanda hala eski yöntemlerle yapılmaktadır. Bu konvansiyonel sınıflandırma yöntemleri zaman israfına ve sınıflandırmada insan kaynaklı hatalara yol açmaktadır. Son yıllarda tarımsal ürün tespiti ve sınıflandırmalarında yapay zekâ çalışmalarının arttığı gözlemlenmektedir. Bu çalışmada da EfficientNet ve ResNet50 derin öğrenme modellerinin limon görüntülerinin sınıflandırılmasındaki başarısı araştırılmıştır. Çalışmada Majiec Adamiak tarafından oluşturulmuş Limon kalite kontrol veri seti (Lemons quality control dataset) adlı kullanılmıştır. Modellerin eğitilmesinde veri çoğaltma (data augmentation) ve öğrenimin aktarılması (transfer learning) işlemlerinden faydalanılmıştır. Basit rastgele örnekleme sonuçlarına göre 30 kez çalıştırılan EfficientNet modelinde başarı ortalaması %69.44 ve kayıp değeri de 0.8104 olarak hesaplanmıştır. ResNet50 modelinde elimizdeki verilerle yapılan beş katlı çapraz doğrulama (5-fold cross validation) sonucunda en yüksek başarı değeri %99.79 olurken ortalama başarıda %98.84 değeri elde edilmiştir. Kayıp değer ortalaması da 0.063 olarak hesaplanmıştır. Modellerin daha önce görmediği limon görüntüleri üzerine yaptığı tahminlerin genel olarak başarılı olduğu görülmüştür. Bazı küflü limonların renginden dolayı yeşil limon olarak algılandığı gözlemlenmiştir. Bu sıkıntının eldeki veri miktarının artırılması ile çözülebileceği düşünülmektedir.
Our country is among the leading countries in the world in lemon production. According to research, one fourth of the world's lemon exports are made by Türkiye. In our country, which has such a high production capacity, lemon classification is still done with old methods in many areas. These conventional classification methods lead to waste of time and human-induced errors in classification. In recent years, it has been observed that artificial intelligence studies have increased in agricultural product detection and classification. In this study, the success of EfficientNet and ResNet50 deep learning models in classifying lemon images was investigated. In the study, Lemons quality control dataset, which was created by Majiec Adamiak, was used. Data augmentation and transfer learning processes were used in training the models. According to the results of simple random sampling, the average of success in the EfficientNet model, which was run 30 times, was calculated as 69.44% and the loss value was calculated as 0.8104. As a result of the five-fold cross validation (5-fold cross validation) performed with the data we have in the ResNet50 model, the highest success value was 99.79%, while the average success rate was 98.84%. The mean lost value was also calculated as 0.063. It was seen that the predictions made by the models on the lemon images that they had not seen before were generally successful. It has been observed that some moldy lemons are perceived as green lemons due to their color. It is thought that this problem can be solved by increasing the amount of data available.
Our country is among the leading countries in the world in lemon production. According to research, one fourth of the world's lemon exports are made by Türkiye. In our country, which has such a high production capacity, lemon classification is still done with old methods in many areas. These conventional classification methods lead to waste of time and human-induced errors in classification. In recent years, it has been observed that artificial intelligence studies have increased in agricultural product detection and classification. In this study, the success of EfficientNet and ResNet50 deep learning models in classifying lemon images was investigated. In the study, Lemons quality control dataset, which was created by Majiec Adamiak, was used. Data augmentation and transfer learning processes were used in training the models. According to the results of simple random sampling, the average of success in the EfficientNet model, which was run 30 times, was calculated as 69.44% and the loss value was calculated as 0.8104. As a result of the five-fold cross validation (5-fold cross validation) performed with the data we have in the ResNet50 model, the highest success value was 99.79%, while the average success rate was 98.84%. The mean lost value was also calculated as 0.063. It was seen that the predictions made by the models on the lemon images that they had not seen before were generally successful. It has been observed that some moldy lemons are perceived as green lemons due to their color. It is thought that this problem can be solved by increasing the amount of data available.
Description
Keywords
Mühendislik Bilimleri, Engineering Sciences
Turkish CoHE Thesis Center URL
WoS Q
Scopus Q
Source
Volume
Issue
Start Page
End Page
65