Optimizing Paper Waste Management with Hyperspectral Imaging
Waste management in general is very important in order to keep economies sustainable and recycling of all sorts are vital in preserving the environment. This research project concentrated on waste paper sorting. Currently paper waste is mainly sorted manually, even though sorting paper is a multi-step process and at certain stages different automated sorting technologies are also being used, hand selection still is more decisive. Automated sorting of waste paper would increase the efficiency, speed and quality of the output streams. The study scope is to discover whether a hyperspectral snapshot camera can be used in automated paper sorting and in final quality control.
The goals of this project were to detect different type of waste paper mainly in a static pile for quality inspection and try to apply it also on a moving conveyor. The main focus was on detecting brownish long-fibered material and distinguishing between wanted and unwanted paper groups, based on 15 different mills prevailing rates. White fibered paper has the highest value on the recycling market and widest use possibility, and it is vital to lower the brownish paper ratio appearing in the paper stream.
Identification of different paper sorts
The research consisted of various measurements at changing conditions to collect and build a database of paper materials and their spectral response as hypercubes, analyze these to what information they contain and propose a classification model.
The experiments were conducted on several samples of paper material, on eight subcategories (Book cover, Brown envelope, Brown and Grey packaging as unwanted class, and Magazines, Newspaper, Office paper as wanted category and background), using Cubert’s hyperspectral camera S185 with a wavelength range between 450 – 950 nm, a spatial resolution of 1000 x 1000 pixels and 125 spectral band.
The analyzes of the acquired hypercubes showed that the 450-650 nm range includes the relevant information for material classification, and could successfully be used for material color detection. Additionally, the greatest distinctive sign could be noted between the brown fibered materials, following a monotonous curve and magazines, newspaper and office paper, following a saturation curve, in addition to the luminosity disparities (Figure 2).
Figure 2.: Reflectance values of the main samples of paper material
For classification, perClass Mira was used, which represents a machine learning algorithm for multi-class classification, based on trainings data. The algorithm is an iterative procedure, reducing the error by iteration, using vectors and their weights.
In total, the proposed model enables to differentiate between eight classes: Book cover, Brown envelope, Brown and Grey packaging, Magazines, Newspaper, Office paper and Background. The classification process consists of various measurements considering changing conditions, like illumination on-site, in lab and outside, in order to collect reasonable amount of training data with a high intensity of spectra on each sample.
For the validation of the classification multiple hypercubes are used, different from the training data and all selected randomly. The highest accuracies are achieved while classifying brown packaging, envelope, book covers and newspaper (on lab data: 90.5%, 95.6%, 70.8%, and 43.3 %; outside data: 47.7%, 60.5%, 34.4% and 57.7%) and the lowest were grey packaging, magazine and office paper (lab data: 8.7%, 26.9% and 41.4%; outside data: 36.8%, 29.2% and 35.9%).
Figure 2 shows that the classification works successfully detecting brown cardboard and carton or brown envelope.
Figure 3: Classified images (from top left: grey packaging, brown package, grey p., office paper, newspaper, small pieces of paper material, brown p., piles of paper material) Legend: grey package (red), newspaper (green), brown package (brown) ,office paper (white), magazine (blue), background (black).