A Novel Machine Vision Based Seed Quality Sorting System: Toward the Industrialization

Seed quality is the main factor affecting the seed germination and thus crop yield. In the course of harvest, seeds normally contain impurities, which are essentially foreign materials, broken seeds, and infected seeds. Including such impurities in seed lot obviously lead to the lowcrop yield. Knowledge of seed lot quality before sowing is important to farmers for yield prediction, and to seed companies for warrant determination. Recently, some laboratory based studies were conducted for rapid detection of quality seeds where few images of limited number of seeds were collected and analyzed with image processing techniques. However, such previous studies majorly sought to development of image processing algorithm for quality seed detection, and only a few presented prototype system for the same purpose. In this work, moving a step forward, a machine vision system has developed for real-time quality screening of tomato and cabbage seed sample. The machine vision system comprises of a color camera, illumination source, conveyor belt, motion controller, a computer unit, and a custom-built software interface for real time image processing and visualization of detected foreign materials and bad seeds. A total of >50 morphological, color, and textural features were firstly extracted from each collected sample (good/bad seed and different foreign materials) images. For the sake of fast computation, optimal features were selected based on random forest and sequential forward selection methods. A multivariate classification method of one-class partial least square (OCPLS) classifier was executed on selected feature to classify imaged samples into two groups as good seed and bad seeds/foreign materials. The developed image processing algorithms and classifiers were incorporated to the software interface for real-time quality seed screening while samples moving through the conveyor unit. The obtained classification accuracy was >95% for tomato seeds and >90% for carrot seeds. The developed machine vision can collect the seed sample images from a 30 cm2 area and the quality screening results can be depicted within few seconds (<3s). Since the conveyor unit was synchronized with camera exposure and image processing, thus can be run gradually for the real-time screening. The proposed machine vision system can be adopted by seed processing companies and large-scale farmers for seed quality screening to improve germination rate of seed lot, crop yield and profitability.