Colombian fruit and vegetables recognition using convolutional neural networks and transfer learning

Abstract

Automatic image recognition is a convenient option for labeling and categorizing fruits and vegetables in supermarkets. This paper proposes the design and implementation of an automatic classification system for Colombian fruits, by training a convolutional neural network. A database was created to train and test the system, which consisted of 4980 images, labeled in 22 classes, each corresponding to pictures of the same kind of fruit, trying to reproduce the variability of a real case scenario with occlusions, different positions, rotations, lightings, colors, etc., and the use of bags. On-training data augmentation was used to further increase the robustness of the model. Additionally, transfer learning was implemented by taking the parameters of a pretrained model used for fruit classification as the new initial parameters of the proposed convolutional network, achieving an increase of the classification accuracy compared with the same model when trained with random initial weights. The final classification accuracy of the network was 98.12% which matches the scores achieved on previous works that performed fruit classification on less challenging datasets. Considering top-3 classification we report an accuracy of 99.95%.

Convolutional Neural Network in Image Recognition

“…(2) Convolutional layer The convolutional computing layer is the most important layer in the CNN structure, and the convolutional layer contains a different number of convolutional filters (also known as convolution kernels), which can extract different features of the input image data [14]. The calculation process of the convolution operation is shown in Equation (1), where L represents the number of layers of the neural network, K represents the convolution kernel, Mj represents the JTH feature map of channel M, I represents the ith feature in Mj, and B represents the bias.…”