Computer vision classification of dry beans (Phaseolus vulgaris L.) based on deep transfer learning techniques
Yavuz Selim Taspinar,Musa Dogan,Ilkay Cinar,Ramazan Kursun,Ilker Ali Ozkan,Murat Koklu
DOI: https://doi.org/10.1007/s00217-022-04080-1
IF: 3.3
2022-08-06
European Food Research and Technology
Abstract:There are many types of haricot beans, the nutrient consumed all over the world. Each type differs in terms of features such as taste, size, economic value, etc. But even if they are different types, bean grains are frequently confused with each other. For these reasons, it is important to separate the bean grains of different species. For this purpose, a haricot bean dataset consisting of 33,064 images of 14 different bean types was created. By using these images, 3 different pre-trained Convolutional Neural Networks (CNN) were trained via the transfer learning method. Within the scope of the study, InceptionV3, VGG16, and VGG19 CNN models were used. These models were utilized for both end-to-end classification and extraction of image features. Firstly, the images were classified via Inception V3, VGG16, and VGG19 models. As a result of this classification, 84.48%, 80.63%, and 81.03% classification success were obtained from InceptionV3, VGG16, and VGG19 models, respectively. Secondly, the image features of these 3 models were taken from the layer just before the classification layer. Then, these features were given as input to the Support Vector Machine (SVM) and Logistic Regression (LR) models. Images were classified using six different models, InceptionV3 + SVM, VGG16 + SVM, VGG1 + SVM and InceptionV3 + LR, VGG16 + LR, VGG1 + LR. Classification successes obtained from InceptionV3 + SVM, VGG16 + SVM, and VGG19 + SVM were 79.60%, 81.97%, 80.64%, respectively. And, the classification successes obtained from InceptionV3 + LR, VGG16 + LR, and VGG19 + LR were 82.35%, 83.71%, and 83.54%, respectively. The InceptionV3, among all models, was determined to be the best classification model with a classification success of 84.48%. On the other hand, the model with the lowest classification success was determined to be the InceptionV3 + SVM. Detailed analysis of the created models was also carried out with precision, recall, and F-1 score metrics. It is thought that the proposed models can be used to distinguish haricot bean types in a quick and accurate way. Furthermore, the proposed computer vision methods can be combined with robotic systems and used to the distinction of bean types. By means of image processing, varieties can be determined on conveyor belts, and dry bean varieties can be purified with delta robots.
food science & technology