Image capturing, segmentation and data analysis of shredded refuse streams.

Refuse sorting is an important cornerstone of the recycling industry, but ever-changing refuse compositions and the desire to increase recycling rates still pose many unsolved challenges. The digitalisation of refuse sorting plants promises to overcome these challenges by optimising and automatically adapting the sorting process. This publication describes a system for image capturing, segmentation-based refuse recognition and data analysis of shredded refuse streams. The image capturing collects multispectral 2D and 3D images of the refuse streams on conveyor belts. The image recognition performs a semantic segmentation of the images to determine the refuse composition from the 2D images, whereas the 3D images approximate the volumes on the conveyor belts. The semantic segmentation is done by a combined convolutional neural network model, consisting of a foreground-background and a refuse class segmentation. Both models rely on synthetic training data to reduce the necessary amount of manually labelled training data, whereas the final segmentation performance reaches an Intersection over Union of up to 75%. The results of the semantic segmentation and volume estimation are combined with data of the shredding machinery by transforming it into a unified representation. This combined dataset is the basis for estimating the processed refuse masses from the semantic segmentation and volume estimation.

Deep Metric Learning with BIER: Boosting Independent Embeddings Robustly.

Learning similarity functions between image pairs with deep neural networks yields highly correlated activations of embeddings. In this work, we show how to improve the robustness of such embeddings by exploiting the independence within ensembles. To this end, we divide the last embedding layer of a deep network into an embedding ensemble and formulate the task of training this ensemble as an online gradient boosting problem. Each learner receives a reweighted training sample from the previous learners. Further, we propose two loss functions which increase the diversity in our ensemble. These loss functions can be applied either for weight initialization or during training. Together, our contributions leverage large embedding sizes more effectively by significantly reducing correlation of the embedding and consequently increase retrieval accuracy of the embedding. Our method works with any differentiable loss function and does not introduce any additional parameters during test time. We evaluate our metric learning method on image retrieval tasks and show that it improves over state-of-the-art methods on the CUB-200-2011, Cars-196, Stanford Online Products, In-Shop Clothes Retrieval and VehicleID datasets. Therefore, our findings suggest that by dividing deep networks at the end into several smaller and diverse networks, we can significantly reduce overfitting.