Apple surface defect detection based on improved CNN and data augmentation

Abstract

Objective: Improve the detection accuracy and efficiency of apple surface defects. Methods: An detection method for apple surface defects was established based on an improved convolutional neural network (CNN) and data augmentation method. Firstly, the classical CNN was improved to detect apple surface defects. Then, using the conditional generation adversarial network, the image data of surface defect free and defective apples was augmented with synthetic apple images to improve the detection performance of the improved CNN for apple surface defects. Finally, by pruning the CNN model, the detection accuracy, detection speed and energy saving limits of apple surface defects were balanced reasonably to improve the practicability of the proposed method. Results: When 2 048 interpretive neurons were selected in the interpretation layer of the improved CNN, the average detection accuracy was the highest among the interpretive neuron number situations. Additionally, the diversity of the apple image data sets was enhanced with the synthetic apple images produced by the conditional generation adversarial network. In addition, the accuracy of the proposed method for detecting apple surface defects increased continuously with the increase of the proportion of the enhanced images in the test data set. When the ratio of the pruned model size to the original model size decreased from 100% to 50%, the detection efficiency of apple surface defects was doubled with 6.96% detection accuracy decreasing. Conclusion: This method is expected to realize the automatic defect detection in apple production and processing, and provide a reference for the developing of other fruit surface defect detection methods.

Keywords

• surface defect
• apple
• improved cnn
• data augmentation