基于深度学习的大米垩白分割算法研究

Research on Chalkiness Rice Segmentation Algorithm Based on Deep Learning

大米质量安全关系了人们的生命健康,而含有垩白的大米因为缺少有助于人体代谢的成份,营养价值低,如何准确快速地检测出大米中的垩白信息就显得尤为重要。本研究提出了一个轻量级的语义分割网络IMUN,该网络由非对称型的编码与解码结构组成。编码结构基于改进的MobileNetV2,使用深度可分离空洞卷积,扩大感受野的同时,能获取更多特征信息。解码结构基于Unet的解码结构,将上采样过程中恢复的特征,与同层编码结构进行特征连接,有助于保留更多细节信息。该网络结构可以实现对大米上的垩白区域的像素级分割,继而可以获取大米的垩白粒率和垩白度。实验结果表明,大米上的垩白区域的分割准确率可达到94.11%,在像素精度和交并比等方面优于FCN及大部分近年来新提出的网络。同时本网络结构参数少,网络模型小,非常适合于集成到嵌入式可移动的检测设备中。

Rice quality is significanly influential to human health,thus it is crucial to detect the chalky rice accurately and fast.A lightweight semantic segmentation network named IMUN,it was proposed to measure chalkiness of rice in this paper.The IMUN consists of an asymmetric encoding and decoding structure.The encoding part was improved by MobileNetV2,uses the deep separable hole convolution,which could expand receptive filed and thus obtain more feature information.The decoding part was based on Unet,and it was connected the features recovered during the upsampling process with the encoding structure to retain more detailed information.The network we proposed can obtain pixel-level segmentation of the chalk rice.In addition,higher accuracy of chalk rice rate and chalkiness degree can be obtainedwhen the network was used in the detection of the chalk rice.The results show that the segmentation accuracy of the chalky block can reach 94.11%,which is better than FCN and most of semantic segmentation networks reported recentlywith respect to pixel accuracy and intersection over union.At the same time,the network structure has fewer parameters and smaller model size.Therefore,IMUN can be easily transplanted into mobile and embedded device.