High-throughput maize grain type identification system based on sparse representation algorithm
Ma Qin,Wang Yue,Guo Hao,Zhu Dehai,Liu Zhe,Zhang Xiaodong,Li Shaoming
DOI: https://doi.org/10.11975/j.issn.1002-6819.2017.z1.033
2017-01-01
Abstract:Maize grain type is one of important phenotype parameters to evaluate maize yield and quality. In order to improve the recognition rate of the maize grain type and achieve high throughput and nondestructive measurement, the maize grain type identification system based on sparse representation is established. For the maize ear's irregular shape and uneven growth, the phenotypic trait acquisition of ear needs to meet the all-dimensional requirements. The hardware acquisition equipment is designed to capture vertically dropping maize ear by 3 high-speed cameras crossed with an angle of 120° mutually in black box. The ear falls down very fast during the process, and thus the high speed of camera's shutter is needed. In addition, enough supplemental lighting is essential because of the high speed of camera. The size of black box is 800 mm × 800 mm × 350 mm. The high-speed CCD (charge coupled device) camera model is DH-SV2001GC, and the image resolution is 1628×1236 pixels. The grain images of 3 varieties i.e. flint grains, dent grains and half-dent grains are taken as the research objects. Firstly frame difference method is used to acquire maize contour, and then G channel separation, median filter and Otsu algorithm are used to segment grain contour. Use concave points matching algorithm to solve grain adhesion problem. Then the color feature parameters (average of L-channel, average of a-channel, average of b-channel), the shape feature parameters (cross sectional area, round degree, elongation, rectangular degree) and the texture feature parameters (angular second moment, contrast, inverse difference moment, entropy, correlation) are extrated, which can distinguish different types of grain as the typical characteristics. A total of 200 grains for each grain type is randomly selected to form the dictionary of the sparse representation method. After that, normalize the over-complete dictionary and every test sample. For each test sample, calculate the sparse representation coefficient, and then determine grain type according to the minimum reconstruction error. The classification algorithm is tested by computer which is configured as Intel(R) Core(TM) i7-4710MQ CPU @2.50GHz and the RAM (random access memory) is 8 GB. The test code is written by C++ and the IDE (integrated development environment) is Visual Studio 2013. The image processing library is OpenCV 2.4.9 and the compressed sensing library is KL1p. Experimental results show that the identification accuracy of that algorithm for the maize grains is 94.8%, and the Kappa coefficient of confusion matrix is 0.923, obtaining a high-level discriminant consistency. The recognition accuracy of half-dent grain is not as high as flint grain and dent grain, and because half-dent grain is in the intermediate state between flint grain and dent grain, the difference between it and the other 2 types is not obvious and false recognition maybe occurs. Experiment shows the measurement speed is up to 28 spikes per minute, which meets the demand of high throughput variety test. So the maize grain type identification system proposed in this paper provides an important technique and method for maize variety test and automatic breeding.