Automated 3D Surface Reconstruction and Analysis of apple Near-Infrared data for the application of apple stem-end/calyx identification

Machine vision technologies have received more and more attention for automated apple sorting and grading applications. Currently most machine vision techniques used in apple automated processing are 2-D or 2.5-D based approaches, which usually have difficulties to identify apple stemend/calyx from defects. On the other hand, as one of the persistent problems, apple stem-end/calyx identification must be solved not only for the accurate defect inspection, but for efficient firmness measurement purposes. To solve above problem, a novel 3-D based apple near-infrared (NIR) data analysis strategy was proposed so that the apple stem-end/calyx could be identified, and hence differentiated from defect and normal tissue according to their different 3D shapes. Two automated 3-D data processing approaches were developed in this research: 1). 3-D quadratic facet model fitting, which employed a small concaved 3-D patch to fit 3-D apple surface, and the best fit could be found around stemend/calyx area; and 2). 3-D shape enhanced transform (SET), which enhanced the apple stemend/calyx area, and made it easily detectable, according to the 3D surface gradient difference between the stem-end/calyx and the apple surface. Apple 3-D surface was reconstructed from 2-D NIR images according to Shape-From-Shading (SFS) method. Unlike 2.5-D approach, the SFS took advantage of the whole image information, which meant a more detailed apple 3D description could be obtained. The proposed 3-D approaches didn't depend on the location of the stem-end/calyx on the apple surface, making it more suitable for apples orientated randomly. There was also no additional light source required in the imaging system: normal visible white light plus a NIR filter was enough. In addition, the proposed 3-D approaches were found to be robust to the noise and incomplete image data. Finally, the experiment results demonstrated the effectiveness of proposed 3-D approaches, and an overall detection rate above 90% was achieved.