Pickled and dried mustard foreign matter detection using multispectral imaging system based on single shot method.

This paper investigated a usage of multispectral imaging system based on single shot method for detecting the foreign matter (FM) in the pickled and dried mustard (PDM) on a moving conveyor belt. Multispectral images of PDM and FM in a quiescent state and the PDM mixed with FM in a moving state were respectively obtained using the multispectral imaging system with a spectral range from 676 to 952 nm and spatial resolution of 409 × 216 pixels. Pure pixel data of PDM and FM were extracted from multispectral images of the PDM and the FM in a quiescent state. For the pixel-level classification, the support vector machine (SVM) and the back propagation neural network (BPNN) were applied to develop models to classify FM and PDM on the full bands, respectively. The classification accuracy and the mean prediction time of SVM model were 98.23% and 6.8s; the classification accuracy and the mean prediction time of BPNN model were 98.07% and 0.04s. The BPNN model was selected as the optimal model considering the classification accuracy and prediction time synthetically. Using the optimal model to detect FM in the PDM during the moving process, the identification accuracy of FM was 97.9%. The results demonstrated that multispectral imaging system could be used for the online detection of foreign matter in the pickled and dried mustard. • SVM and BPNN models are used to classify PDM and FM at the pixel level. • Based on the optimal model, the FM in PDM are detected during the moving process. • The model performance is evaluated using images of the PDM mixed with FM. • The proposed method exhibits good results in detecting the FM in PDM. [ABSTRACT FROM AUTHOR]