Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Ensemble neural network methods for the sorting of recyclables

Vogiatzis Antonios

Full record

---

Summary

The classification of recyclable materials, particularly the recovery of plastic, is critical not only for economic sustainability, but also for environmental sustainability. Deep learning is a machine learning paradigm that employs artificial neural networks with multiple layers to progressively extract higher level features from raw input; a paradigm that in recent years has achieved extraordinary success in a wide variety of applications, including the classification of materials. This thesis introduces two novel ensemble neural network methods for image classification. Our methods build on the concept of “shared wisdom from data” in order to effectively classify recyclable materials. Specifically, in the first part of this thesis we introduce the so-called “Dual-branch Multi-output CNN” based on recent work in deep learning and waste classification. This is a custom convolutional neural network composed of two branches that aims to i) classify recyclables and ii) distinguish the type of plastic. The proposed architecture consists of two classifiers that have been trained on two distinct datasets in order to encode complementary characteristics of recyclable materials. Our approach allows the learning of disjoint label combinations, by making use of the datasets’ joint utilization—but without requiring their mingling. In the second part of this thesis, we propose a generic classification architecture based on independent parallel CNNs that explicitly exploits a “mutual exclusivity” or “classifiers’ mutually supported decisions” property that exists in many dataset domains of interest, namely that an image in a given dataset may almost unquestionably belong to only one class. Our framework incorporates several purpose-built opinion aggregation decision rules that are triggered when the mutual exclusivity property is satisfied or not; and it makes use of “weights” that intuitively reflect each CNN’s confidence in correctly identifying its corresponding class. Thus, our framework can (a) take advantage of clearly defined class boundaries when they exist, and (b) successfully assign items to classes when clearly defined class boundaries do not exist. Our experiments, with two well-known problem-specific image datasets, confirm the effectiveness of our ensemble neural network methods in the classification of recyclables.

Available Files

Services