Extraction of litchi fruit pericarp defect based on a fully convolutional neural network

Objective  To enhance the effects of litchi fruit pericarp defect extraction and satisfy the accuracy requirements of quality detection and classification.

Method  A fully convolutional neural network was built up based on AlexNet (AlexNet-FCN) using Tensorflow framework, with ReLU as the activation function, Max-pooling as the down-sampling method and loss function of Softmax regression classifier as the optimization target. Mini-batch stochastic gradient descent (Mini-batch SGD) was used to optimize the model.

Result  When the model was converged, the intersection-over-union of dehiscent area (IoU_d) of litchi fruit cracking was 0.83 for the validation set, the intersection-over-union of brown area (IoU_b) was 0.60, and the intersection-over-union of both dehiscent and brown area (IoU_a) was 0.68. Compared with linear-support vector machine (SVM) and Naïve Bayes classifier, AlexNet-FCN had a stronger defect extraction ability.

Conclusion  Fully convolutional networks (FCN) have a good prospect for application of fruit pericarp defect extraction.