Transfer Learning with Efficient Convolutional Neural Networks for Fruit Recognition

An efficient and effective image based fruit recognition network is critical for supporting mobile application in reality. This paper presents a method to recognize fruit faster and more accurately by using the transfer learning technique. The proposed network performs depthwise separable convolution with thinner factor to reduce the size of vanilla network and improve the performance by adapting global depthwise convolution. Additionally, we make a simple analysis on how those methods reduce the parameters and the cost of computation in training process. In order to test the accuracy and enhance the robustness of the model, we use Fruits-360 dataset which contains 55244 images spread across 81 classes. The experimental results demonstrate that our proposed network is superior to three previous state-of-the-ad networks. Moreover, our model has a higher accuracy than the vanilla model with the same thinner factor.