Novel Statistical Pattern Recognition and 3D Machine Vision Technologies for Automated Food Quality Inspection

Abstract

Machine vision technologies have received a lot of attention for automated food quality inspection. This dissertation describes three techniques developed to improve the quality inspection of apple and poultry products. First, a Gabor feature-based kernel principal component analysis (PCA) method was introduced by combining Gabor wavelet representation of apple images and the kernel PCA method for apple quality inspection using near-infrared (NIR) imaging. Gabor wavelet decomposition was employed to extract appropriate Gabor features of whole apple NIR images. Then, the kernel PCA method with polynomial kernels was applied in the Gabor feature space to handle nonlinear separable features. The experimental results showed the effectiveness of the Gabor-based kernel PCA method. Using the proposed Gabor kernel PCA eliminated the need for local feature segmentation and also resolved the nonlinear separable problem in the Gabor feature space. An overall 90.5% detection rate was achieved. Second, a novel 3D-based apple near-infrared (NIR) data analysis strategy was utilized so that the apple stem-end/calyx could be identified, and hence differentiated from defects and normal tissue according to their different 3D shapes. Two automated 3D data processing approaches were developed in this research: 1) A 3D quadratic facet model fitting, which employed a small concave 3D patch to fit the 3D apple surface and the best fit could be found around stem-end/calyx area; and 2) A 3D shape enhanced transform (SET), which enhanced the apple stem-end/calyx area and made it easily detectable because of the 3D surface gradient difference between the stem-end/calyx and the apple surface. An overall 92.6% accuracy was achieved. Third, high resolution on-line laser 3D imaging was investigated for improving the 3D profile recovery for thickness compensation purposes. Parallel processing and memory management were also considered to improve the processing speed of the detection system. Multiple-lane coverage was fulfilled such that a wider conveyor could be used and overall throughput would be increased. To further improve the detection performance of the dual X-ray and laser imaging system, a dynamic thresholding approach was introduced to suppress the errors and noise involved by the imaging system. Unlike the traditional single threshold method, dynamic thresholding monitored the responses of the region of interest under a set of thresholds to determine the true physical contaminants, making it more tolerant to the noise than the single threshold method. An overall 98.6% detection rate was achieved.