Rapid Detection System of Paddy Impurities Based on Machine Vision

Abstract

Aiming at the problems of poor imaging effect and low impurity classification accuracy of existing paddy quality detection systems, we propose a rapid paddy impurity detection system based on machine vision. The system consists of an image acquisition subsystem and an impurity classification subsystem, which aim to accurately evaluate the quality of paddy grains. The image acquisition subsystem is composed of a multi-degree-of-freedom manipulator, a flexible vibrating disc, an industrial camera, a light source, and an embedded computing platform. The multisensor fusion technology is used to realize the two-dimensional imaging of the target paddy. Moreover, the impurity classification subsystem innovatively adopts the feature extraction method based on convolutional neural network to realize the fast and accurate classification of paddy impurity images. To balance the classification accuracy and computational complexity, we use GhostNet as the backbone network for efficient high-dimensional semantic feature extraction from paddy images. Furthermore, the visual attention mechanism and multiscale feature aggregation strategy are introduced to enhance the extracted features and improve the accuracy of impurity classification. The image acquisition subsystem collects 256 raw images with a resolution of 5120 × 5120. After image preprocessing, a total of 5000 paddy image blocks are obtained, and each block has different resolutions. These image blocks include lesion paddy images, moldy paddy images, sprout paddy images, immature paddy images, and normal paddy images. The trained impurity classification model achieves 83.64% precision, 83.60% recall, 83.54% F1 score, and 93.44% accuracy. These results not only show the efficiency and accuracy of the system in practical applications but also provide important reference value for future research and application in a wider range of agricultural detection fields.