Anomaly Detection in Fruits using Hyper Spectral Images

Abstract

One of the biggest problems in hyper spectral image analysis is the wavelength selection because of the immense amount of hypercube data. In this paper, we introduce an approach to find out the optimal wavelength selection in predicting the quality of the fruit. Hyper spectral imaging was built with spectral region of 400nm to 1000nm for fruit defect detection. For image acquisition, we used fluorescent light as the light source. Analysis was performed in visible region, which had spectral from 413nm to 642nm it was done because of the low reflectance spectrum found in fluorescent light sources. The captured image in this experiment demonstrates irregular illumination that means half of the fruit has brighter area. Analysis of the hyper spectral image was done in order to select diverse wavelengths that could possibly be used in multispectral imaging system. Selected wavelengths were used to create a separate image and each image went through thresholding. Experiment shows a multispectral imaging system which is able to detect defects in fruits by selecting most contributing wavelengths from the hyper spectral image. Algorithm presented in this paper could be improved with morphology operations so that we could get the actual size of the defect.

Power spectral density-based fractal analyses of sputtered yttria-stabilized zirconia thin films

“…The PSD data were then imported to ORIGIN 2018 software to smoothing. We applied a low-level Savitzky-Golay filter to enhance the signal-to-noise ratio without extremely signal distortion [33]. It is also called a digital smoothing polynomial filter or a least-squares smoothing filter.…”

Thoughts for Foods: Imaging Technology Opportunities for Monitoring and Measuring Food Quality

“…Therefore, for example with potatoes or apples, there is a risk of having readings that can also depend on the variable distance from the emitted source of light. For this reason, images should be observed and sampled in areas where the distance from the sensor is equal for all samples [9]. So, the first step to take is to cut out the sub-sample of photography or better than more sub-photography samples in which to read the amount read by the sensor.…”