Computer Vision
We develop solutions based on computer vision and deep learning for any industry. We accompany you throughout the process, designing a tailored solution.
Objective. Automatically and accurately classification of plastic waste with hyperspectral.
The ADITIV-SOSTENIBLE (PHASE II) project, “Research of technologies for flexible and intelligent additive manufacturing processes of sustainable materials” has the overall objective of researching and developing technologies for additive manufacturing, using robotics, intelligent monitoring systems, and sustainable materials. Within this project, we specifically explain one of its lines of work: the development of an automatic plastic waste classification system based on hyperspectral cameras. The objective of this specific part is to automatically and accurately classify different types of plastics present in waste, with special attention to those generated in the industry of Aragón.
This project has received public funding under the AEI 2023 call from the Ministerio de Industria, Comercio y Turismo, within the established aid for the support of innovative business clusters in order to improve the competitiveness of small and medium-sized enterprises, with the support of the European Union through the Next Generation EU program, within the Plan de Recuperación, Transformación y Resiliencia.
The total budget of the project is €287,344, with a grant awarded of €221,528.
The participating partners in ADITIV-SOSTENIBLE (PHASE II) are: the Clúster de Automoción de Aragón (CAAR), ASAI Industrial, INFINITIA, the Instituto Tecnológico de Aragón (ITAINNOVA) and the Fundación Clúster de Empresas de Automoción de Galicia (CEAGA).
Solution. We developed a classification system using hyperspectral cameras.
ATRIA’s role has focused on the classification of plastic waste process with hyperespectral, to the development of automatic classification systems based on artificial intelligence. Our approach focuses on providing an efficient solution for waste separation and plastic valorization.
First, we obtained representative samples of plastic waste from various industries. For polymer identification, we used NIR (near-infrared) hyperspectral cameras. Unlike conventional cameras, which only capture information in three color bands (RGB), these record spectral reflectance in hundreds of narrow bands within the NIR spectrum. This provides us with a unique and detailed spectral “fingerprint” for each type of material, essential for accurate plastic classification. We took images of a wide variety of common polymers (such as PE, PP, PET, PVC, ABS, LDPE, PPH, and PC), both of individual samples to establish a robust spectral database and of complex mixtures to simulate real waste sorting scenarios.
We digitally processed the hyperspectral images to improve their quality and remove noise, applying spectral analysis and image processing techniques. We then developed and trained various classification models based on machine learning algorithms. These computer vision algorithms “learn” to recognize and distinguish the different types of plastic from their unique spectra. To optimize the system’s performance and speed, we implemented band selection techniques, allowing the system to focus only on the most relevant spectral information for polymer classification, making the process more efficient and suitable for high-speed recycling lines.
The results demonstrate the high accuracy of our system in classifying plastic waste. We achieved correct classification in a very high percentage of cases, both with individual polymer samples and with mixtures of plastic waste, validating the robustness and reliability of the solution for industrial applications of circular economy and waste management.
