基于生成对抗网络和Mask R-CNN的苹果早期变质检测

Early spoilage detection of apple based on generative adversarial network and Mask R-CNN

[目的]提高苹果早期变质区的检测准确率。[方法]基于生成对抗网络和卷积神经网络技术的苹果变质区检测方法。利用Pix2PixHD模型生成包含采后早期变质区的贮藏苹果的近红外成像数据;使用Mask R-CNN模型对生成的近红外图像进行分割,以检测苹果中的变质区;在具有人工智能功能的低成本嵌入式系统上,利用生成的近红外成像数据,实施基于生成对抗网络和卷积神经网络技术的采后苹果的早期变质区域分割和预测。[结果]该方法对收获后苹果的早期变质检测平均准确率比其他9种方法高1.825%~10.435%;Pix2PixHD能以17帧/s的速度从RGB图像生成了可视近红外图像,Mask R-CNN能够以4.2帧/s的速度对苹果图像中的变质区域进行分割。[结论]研究提出的方法有望促进低成本食品质量控制器的开发。

[Objective]To improve the detection accuracy of early apple spoilage zone.[Methods]An apple spoilage detection method was proposed based on generative adversarial network and convolutional neural network.The Pix2PixHD model was used to generate near-infrared imaging data of stored apples in the early postharvest metamorphic area.The Mask R-CNN model was used to segment the generated near Infrared image to detect the deterioration zone in the apple.Based on generative adversarial network and convolutional neural network technology,the early deterioration region segmentation and prediction of postharvest apples were implemented by using the generated near-infrared imaging data on a low-cost embedded system with artificial intelligence function.[Results]The average accuracy of this method was 1.825%~10.435%higher than that of the other nine methods.The Pix2PixHD generated a visible NIR image from an RGB image at 17 frames per second,and the Mask R-CNN was able to segment spoilage areas in an apple image at 4.2 frames per second.[Conclusion]The proposed method is expected to facilitate the development of low-cost food quality controllers.