Estimation on Hole Seeding Quantity of Super Hybrid Rice Based on Machine Vision and BP Neural Network
Tan Suiyan,Ma Xu,Wu Lulu,Li Zehua,Liang Zhongwei
DOI: https://doi.org/10.3969/j.issn.1002-6819.2014.21.024
2014-01-01
Abstract:Super hybrid rice that is widely cultivated in China is mainly planted by transplanting techniques. In super hybrid rice transplanting, a seedling nursery is a preliminary but critical one. Complete rice seedling nursing operations include seed selection, seed germination, sowing of seeds onto seeded trays, and seedling hardening. The sowing process is performed on an automated seeder sowing test line. Matured seedlings are transplanted to the field by transplanters, which are specially designed for transplanting rice seedlings.In practice, some problems arise during the sowing process. Because super hybrid rice has strong tiller ability, it requires a precise and low seeding rate, which needs to ensure 2-3 grains in each cell in tray plugs. Furthermore, rice seed traits, such as length, shape, moisture, and weight change during the sowing process, which greatly affects the performance of the seeder sowing machine. As a result, seed distribution on the tray plugs is uneven, and cavity and single grain rates in tray plug cells are high, which is up to 20%. The low seeding rate will lead to a low seedling survival rate. Also, the estimation accuracy of grains that are overlapped and adhered is quite low. To solve the problem, it is necessary to add nursery cell tray sowing quantity estimation to an automated rice sowing test line. When cavities and single grains in tray cells are detected, artificial reseeding and working parameter adjustment to the seeder sowing machine are needed.After rice seeds are sowed onto the nursery trays, seed connected regions extracted from the acquired image may occur as the following situations: impurity, single grain, and grains that are overlapped and adhered. To a great extent, the shape features of each connected region can determine the grain quantity. In this paper, a method was presented to estimate the sowing quantity per cell in the tray plug, based on machine vision and a BP neural network. The method consisted of four stages:image acquisition and preprocessing, target detection area and cell plug located, seed connected region feature extraction, and BP neural network classification. First, a series of pretreatments to the image were applied, including smoothing, denoising, image binarization, and segmentation. Second, a map projection method of a binary image was performed to locate the target detection area and cell plug. Third, three shape geometric features and seven invariant moments of each seed connected regions were extracted, which were used for inputs of a BP neural network and grain quantity estimation. And last, a BP neural network classifier was employed to classify seed connected regions into impurities, one particle, two particles, three particles, four particles, or five particles and above. After BP neural network classification was done, the seedling tray sowing quantity of each plug cell can be obtained. The result showed that the overall accuracy of the BP neural network was up to 94.4% and the classification accuracy of impurity, one particle, two particles, three particles, four particles, five particles and above were 96.6%, 99.8%, 97.2%, 92.5%, 86.0%, and 94.3%respectively. Each image was processed less than 0.823 seconds. It’s proved that the method of nursery cell tray sowing quantity estimation meets the requirement of an automated rice sowing test line.