Özet:

Automatic quality control of food and agricultural products is a difficult problem that is needed to determine the product quality accurately, quickly and objectively due to the high quality and safety standards of the industry. The solution to this problem requires the automatic classification of products before quality control. In this study presented on the basis of this need, the focus was on the problem of the automatic classification of fruits and vegetables with the help of the machine. To this end, the original and Gauss noise, salt and pepper noise, the disturbances of brightness, and the disturbances of the public data set were applied to the different conditions, including the AlexNet, Vgg16, GoogLeNet, ResNet18, ResNet101, SqueezeNet and ShallowNet popular deep learning architectures. The performance of deep learning architectures in the study was evaluated with accuracy, accuracy, sensitivity, F-dimension and work time with the cross-precision approach in Matlab environment using a personal laptop. Experimental results showed that the network depth is low, the shallowNet with a simple architecture has achieved reasonable achievements and is more vulnerable to disturbances, the number of parameters that can be trained in cases where there are no severe disturbances in the images is low and the SqueezeNet architecture, a deep network, promises hope with a high accuracy level of more than 80%. In addition, the Gauss noise observed that the GoogLeNet architecture, the salt and pepper noise, the Resnet101 architecture, and the brightness changes, the SqueezeNet architecture were more sharp. With a broader set of data that will include more fruit and vegetable categories in the future, repeating the experiments is important and necessary for both testing the accuracy of the above findings and providing the industry with a more realistic machine with a vision solution.

Anahtar Kelimeler: