Anomaly Detection in Polymer Recycling

Description

The "Anomaly Detection for the Singulation of Plastic Wastes in Polymer Recycling" project is part of the practical course Computer Vision for Human-Computer Interaction at Karlsruhe Institute of Technology. It focuses on counting the number of plastic lids in trays and identifying singulation anomalies, addressing a crucial need in polymer recycling processes.

Objective

The project's objective is to develop robust deep learning models that can accurately count and detect anomalies in plastic lid singulation. This is crucial for improving the sorting and recycling of plastics, a significant environmental concern.

Dataset and Data Augmentation

The dataset comprises images of black trays with 0 to 5 plastic lids. Due to class imbalance and limited annotations for instance segmentation, the team innovated a copy & paste data augmentation technique. This method enhanced the dataset with synthetic images, using techniques like object rotation, color changes, and transparency, to improve model training and performance.

Methodology

Two primary approaches were explored:

1. Classification: Using a ResNet18 model trained on both original and synthetic datasets, achieving remarkable accuracy in identifying the number of lids.
2. Instance Segmentation: Employing Mask R-CNN, although it showed lower performance than classification, it provided valuable insights and interpretability.

Results and Future Directions

The classification approach, especially with synthetic data, showed the best performance. Future work could explore combining classification and instance segmentation to enhance both accuracy and interpretability.

The qualitative and quantitative results can be found in the presentation and report in the materials folder.

Installation

Install dependencies using Anaconda.

conda env create -f environment.yml

A GPU unit is required.

Getting Started

For an overview of the project's functionality, see Jupyter Notebooks. The project includes three main modules:

1. Data Augmentation for synthesizing new data.
2. Image Classification using the classification approach.
3. Image Segmentation using instance segmentation.

Before starting, configure the GPU settings in image_segmentation/train_net.py and image_classification/transfer_learning.py. It is recommended to define the store and project directory in image_classification/constants.py.