Striping noise removal method in meat detection based on hyperspectral imaging

Abstract: High-resolution hyperspectral imagery has been one of the most fashionable tools using the push broom scanner in motion, particularly for the in-line measurement in industrial production of food and agricultural products. However, a distinct pattern of stripes can tend to impair the image quality that resulted from the non-uniformity among the sensing pixels, due to the inhomogeneity of imaging sensors in manufacturing, dark currents, and working conditions. Taking the pork as the measurement subject, this work aims to devise a de-stripe procedure for the relative radiance calibration using multiple standard reflectance panels. The specific procedure included the illumination of the standard panels to generate spatially uniform-distributed fluxes on the different levels into the camera, in order to gage the response of individual pixels on the imaging sensor. A polynomial fitting of sensor readouts on reflectance values was implemented to profile the response function of individual sensing pixels, and the compensation of all the sensing pixels for each wavelength, according to the response differences to a designated pixel as reference. A reflectance calibration was also provided to calculate the imaging subject reflectance using this procedure, according to the inverse of the fitted response function for the individual sensing pixels. Furthermore, the hyperspectral images of pork were selected to verify the de-stripe performance of the calibration procedure working on the third-order polynomial fitting with that of moment-matching using an intuitive visual evaluation aided with pseudo-color mapping, an objective profiling of the average digital number per column, as well as the maxima of stripe-index. In addition, a step-reflectance standard panel was used to measure the accuracy of the reflectance calibration on the regions of four reflectance values. Results showed that the subject spatial variation across the image columns was dropped along with the removal of stripe-noise, where the apparent pseudo edges were created, especially at the columns located in the transition from background to subject. There was a significant difference in moment matching in the image-columns average digital numbers of the same gray value after calibration. The calibration per individual pixel working on each wavelength substantially reduced the stripe noise, while no new noise was introduced. The spatial distributions in the original image were perfectly maintained as the image-columns average-digital-number profiles after de-stripe calibration, indicating a good agreement with those before. Aside from the compelling visual evidence, the procedure was also quantified in both the maxima of stripe-index after calibration and the lowness of reflectance error. Specifically, there was a 63.5% decrease in the stripe-index maxima on the hyperspectral images of pork. The reflectance calibration error was 0.003-0.089 for the four regions on the step-standard panel in the most stripe-stricken waveband 200, and 0.009-0.083 in a random waveband 300. The procedure of relative radiance calibration and the companion reflectance calibration can be widely expected to reduce the stripe-noise, while for the intact of subjects' spatial details and precise reflectance values. The finding can be paving the way to better spatial visualization of quality indicators and attributes to the food and agricultural products when working on the images collected using push-broom hyperspectral imagery.