The purpose of this project was to test and implement recent research of polarization and scatter properties that suggest using a cross polarization imaging system to reduce glare artifacts. In particular, the use of this research is to improve the machine vision of apple quality detection in the food industry. The automated measurement system was implemented by acquiring pictures at different angles and different polarization states of apples. The opto-mechanics, system integration, synchronization and data collection are controlled with LabVIEW.
Common illumination systems in short wave infrared (SWIR) hyperspectral imaging (HSI) include direct or
indirect tungsten halogen lights. While direct lights provide more radiation onto the samples than dome setups,
thus being more energy efficient, the acquired images often suffer from specular reflections and gloss. Glare
artifacts in images increase variability in the data limiting the accuracy of machine vision algorithms for defect
detection and quality inspection, or even providing false positives. Although domes are known to provide a
near Lambertian illumination and glare free images, glossy regions and heterogeneities may remain in the data
in practice. More particularly, in the field of fruit and vegetable quality inspection, due to their waxy surface,
it remains challenging to design an efficient realistic lighting system. This paper suggests a new approach
to optimize the illumination of fruit and vegetables based on measurements of the bidirectional reflectance
distribution function (BRDF), shape and Stokes parameters. From these measured values, a BRDF model is
loaded into ray-tracing software for realistic illumination engineering in order to determine the most suitable
illumination scheme. This concept is applied to apples and a cross polarizer (CP) with freeform optics (FO)
optical configuration is proposed, which allows the FO to be optimized to maximize uniformity in the field
of view of the imager and removes the parallel polarized gloss on the apples. The performance of this CP
illumination system was determined experimentally for a set of apples. This cross polarized (CP) illumination
system provided a uniformity (U) of 92% and an efficiency (ν) of 90%, while U = 87% and ν = 14% for an
ideal dome configuration when illuminating a rectangular target. The simulated imaged apples with assigned
optical properties performed better with CP (U=80%) than when using a dome (U=73%) by 7%. Finally, the
sensitivity of the design for the light positioning and spectral tolerance are investigated.
As line scanning short wave infrared (SWIR) hyperspectral imaging (HSI) is a growing field in the food industry, it is important to select efficient illumination designs to image contaminants with high contrast and low noise. Illumination systems can efficiently be compared and optimized through the use of ray tracing simulations. However, these simulations provide illumination patterns in absolute radiometric units while HSI systems typically provide relative measurements. To bridge this gap, a supercontinuum laser and monochromator setup was used in this study to calibrate a SWIR HSI imager in spectral radiometric units. For the radiometric calibration, an integrating sphere (IS) was illuminated with the monochromatic laser light, while both a high sensitivity photodiode and the hyperspectral camera were positioned at different ports of the IS to measure the diffuse light synchronously. For each spectral band, the radiance observed by the imager corresponding to a line was detected using image analysis, while the remainder of the image was used to sample the noise of the sensor. Laser power fluctuations were monitored using a power meter coupled with a thermal sensor, allowing for their correction. As these measurements were time consuming, while InGaAs based sensors are very sensitive to thermal drift, the dark current was sampled frequently to avoid noise time drifts. This approach allowed correcting for 6% of temporal noise fluctuations. A per-pixel linear radiometric model was fitted with an R2 of 0:94±0:3 and used to transfer the measured light distribution of a halogen spot with and without a diffuser into absolute radiometric units. This allowed comparing measurements with the results of ray tracing.
Computer assisted optimal illumination design is crucial when developing cost-effective machine vision systems. Standard local optimization methods, such as downhill simplex optimization (DHSO), often result in an optimal solution that is influenced by the starting point by converging to a local minimum, especially when dealing with high dimensional illumination designs or nonlinear merit spaces. This work presents a novel nonlinear optimization approach, based on design and analysis of computer experiments (DACE). The methodology is first illustrated with a 2D case study of four light sources symmetrically positioned along a fixed arc in order to obtain optimal irradiance uniformity on a flat Lambertian reflecting target at the arc center. The first step consists of choosing angular positions with no overlap between sources using a fast, flexible space filling design. Ray-tracing simulations are then performed at the design points and a merit function is used for each configuration to quantify the homogeneity of the irradiance at the target. The obtained homogeneities at the design points are further used as input to a Gaussian Process (GP), which develops a preliminary distribution for the expected merit space. Global optimization is then performed on the GP more likely providing optimal parameters. Next, the light positioning case study is further investigated by varying the radius of the arc, and by adding two spots symmetrically positioned along an arc diametrically opposed to the first one. The added value of using DACE with regard to the performance in convergence is 6 times faster than the standard simplex method for equal uniformity of 97%. The obtained results were successfully validated experimentally using a short-wavelength infrared (SWIR) hyperspectral imager monitoring a Spectralon panel illuminated by tungsten halogen sources with 10% of relative error.
Visible-near infrared (Vis-NIR) and short wave infrared (SWIR) hyperspectral imaging (HSI) are gaining interest in the food sorting industry. As for traditional machine vision (MV), spectral image registration is an important step which affects the quality of the sorting system. Unfortunately, it currently still remains challenging to accurately register the images acquired with the different imagers as this requires a reference with good contrast over the full spectral range. Therefore, the objective of this work was to develop an accurate high contrast checkerboard over the full spectral range. From the investigated white and dark materials, Teflon and Acktar were found to present very good contrast over the full spectral range from 400 to 2500 nm, with a minimal contrast ratio of 60% in the Vis-NIR and 98 % in the SWIR. The Metal Velvet self-adhesive coating from Acktar was selected as it also provides low specular reflectance. This was taped onto a near-Lambertian polished Teflon plate and one out of two squares were removed after laser cutting the dark coating with an accuracy below 0.1 mm. As standard technologies such as nano-second pulsed lasers generated unwanted damages on both materials, a pulsed femto-second laser setup from Lasea with 60µm accuracy was used to manufacture the checkerboard. This pattern was monitored with an Imec Vis-NIR and a Headwall SWIR HSI pushbroom hyperspectral camera. Good contrast was obtained over the full range of both HSI systems and the estimated effective focal length for the Vis-NIR HSI was determined with computer vision to be 0.5 mm, close to the lens model at high contrast.