Study and Test of the Model of Unsound Kernel Measuring Instrumengt for Distiller Sorghum

Abstract

In view of the low efficiency and low recognition accuracy of sorghum unsound kerneld etection in wine-making enterprises at the present stage, the sorghum unsound kernel detection instrument with image recognition, which was combined with the existing cereal unsound kernel detection instruments in the market, was developed. A series of research was focus on the image collection, key hardware, machine vision and deep learning. In this study, single feature analysis technology, machine-learning based image classification technology, deep-learning based image classification technology and fine-grained image classification technology were used to classify and identify sorghum images, respectively. By contrast, tensorrt deployment technology was used to deploy fine-grained image classification network into the device. The results showed that the recognition accuracy of the developed rapid sorghum imperfect grain detector was less than 1% of the average error of manual detection. The detection time of 50 g sorghum sample was controlled within 5 minutes. Compared with the traditional manual detection, the detection time was greatly shortened, and the subjective deviation of manual detection was also avoided. It is of great significance for the detection and identification of sorghum imperfection rate in wine-making enterprises.

Keywords

• sorghum
• unsound kernel
• deep learning
• fine-grained image classification