Evidential Classification of Incomplete Data via Imprecise Relabelling: Application to Plastic Sorting

Besides ecological issues, the recycling of plastics involves economic incentives that encourage industrial firms to invest in the field. Some of them have focused on the waste sorting phase by designing optical devices able to discriminate on-line between plastic categories. To achieve both ecological and economic objectives, sorting errors must be minimized to avoid serious recycling problems and significant quality degradation of the final recycled product. Even with the most recent acquisition technologies based on spectral imaging, plastic recognition remains a tough task due to the presence of imprecision and uncertainty, e.g. variability in measurement due to atmospheric disturbances, ageing of plastics, black or dark-coloured materials etc. The enhancement of recent sorting techniques based on classification algorithms has led to quite good performance results, however the remaining errors have serious consequences for such applications. In this article, we propose an imprecise classification algorithm to minimize the sorting errors of standard classifiers when dealing with incomplete data, by both integrating the processing of classification doubt and hesitation in the decision process and improving the classification performances. To this end, we propose a relabelling procedure that enables better representation of the imprecision of the learning data, and we introduce the belief functions framework to represent the posterior probability provided by a classifier. Finally, the performances of our approach compared to existing imprecise classifiers is illustrated on the sorting problem of four plastic categories from mid-wavelength infra-red spectra acquired in an industrial context.