Recognition Algorithm of Rice Germ Integrity Base on Improved Inception v3
Bing Li, Shuofeng Li · 2019
This paper adopts a deep learning inceptionv3 as a framework, improves a loss function and adds large boundary regularization parameters, thereby increasing an inter-class difference of germ integrity; and moreover, this paper adds new loss function for feedback, identifies an inter-class boundary of the germ integrity and learns different local features, i.e. different germ integrity features of rice grains on different channels. The original dropout is improved and added to an orthogonal classification layer to reduce the unnecessary link weight of different germ integrity classes, thereby identifying the germ integrity of the rice. The experiment result shows that the proposed algorithm has high identification rate for identifying the integrity of the milled rice with germs, and the comprehensive segmentation accuracy is 94.12%. Compared with other classical identification, the accuracy of the algorithm is greatly improved.