The quality of the kernels in the sunflower seed shell directly affects the quality of the edible oil. Using terahertz time-domain transmission imaging technology combined with morphological filtering and K-means image segmentation, the three abnormalities of sunflower seed like kernel damaged, kernel worm-eaten and empty seed were investigated to achieve to explore the quality of seed kernels in sunflower seed hull. According to the national standards and previous experience, three kinds of shelled sunflower seeds samples of kernel damaged, kernel worm-eaten and empty shells were prepared. The terahertz time domain spectrometer TeraPulse 4000 and transmission imaging accessory were used to acquire the spectral images of the above three abnormal samples at a resolution of 0. 2 mm, and a spectral image of normal sunflower seed was taken as a reference. Four sunflower seed terahertz images were reconstructed by peak-to-peak imaging. The terahertz images can preliminarily determine the shape of the seed kernel in the shell, but there were still problems such as low contrast and blurred edge information, which needed further optimization. The morphological filtering algorithm was used to filter the terahertz images of sunflower seeds. The flat diamond structure element with a side length of 3 was selected as the collation image for one expansion, and then the external gradient of the image was calculated to complete the filtering process of the image. At the same time, the morphological filtering results were compared with the median filtering results, the mean filtering results and the non-local mean filtering results. It was found that the morphological filtering not only ensured the sharpness of the image, but also preserved the edge information, and could also make an obvious boundary between the sunflower seed sample and background, which was conducive to subsequent image segmentation processing. In order to more accurately detect the shape of the sunflower seed kernel, the filtered images were segmented. The K-means clustering algorithm was used to segment the filtered terahertz images. In order to improve the accuracy of the segmentation results, the number K of different initial cluster centers was determined for the images of different samples, among which the kernel damaged K = 4 , the worm-eaten kernel K =5 , the empty shell K = 3, the normal grain K = 4. The segmented images accurately showed the morphology of the kernels in the sunflower seed hull. The study showed that the terahertz time-domain transmission imaging technique combined with morphological filtering and K-means image segmentation method was feasible for the non-destructive detection of the internal quality of sunflower seeds, which laid a foundation for the establishment of non-destructive testing model for the quality of shelled sunflower seeds, and provided a new method reference for nondestructive detection of the internal quality of shelled oil crops.
