Real-Time Detection of Colour and Surface Defects of Maize Kernels Using Machine Vision

Abstract

Real-time feature extraction algorithms were developed using a Matrox IM-1280 real-time image processing board and an Image-RTP real-time processor. On-board hardware operations and parallel statistical look-up table (LUT) mapping were used for high-speed feature extraction. Algorithms were developed to acquire and process geometric features for an artificial intelligence-based automated on-line maize quality inspection system. Maize quality-related features included average colour in red, green, and blue (RGB) and hue, saturation, and intensity (HSI) colour discriminations for specific regions of kernels. Maize quality-related features also included a Fourier profile shape descriptor, which distinguished between shapes of whole and broken kernels. Primitive features can be extracted in a processing time of less than 1 s for one object. With increasing numbers of objects, the processing time needed decreased to 2 s for 12 objects. With proper calibration, primitive feature measurements are faster and can be more accurate than measurements determined by a micrometer. For the RGB colour discrimination, the blue component provides the most separation between the white and yellow maize kernels. The green component provides some separation and the red component provides the least separation between white and yellow maize kernels. For the HSI colour discrimination, the best separation between the white and the yellow maize varieties is provided by the hue component. The processing time was about 1·3 s. Based on the Fourier profile descriptor, the breakage classifier had an accuracy of 95% in classifying whole kernels as whole and 96% in classifying broken kernels as broken. The processing time of the breakage inspection programme was about 1·5 s.