The project "Anomaly Detection for the Singulation of Plastic Wastes in Polymer Recycling" is done in scope of the practical course Computer Vision for Human-Computer Interaction at Karlsruhe Institute of Technology.
The main purpose of the project is counting the number of plastic lids in trays and identification of singulation anomalies. Two main approaches are investigated: image classification and instance segmentation. The performance of both ones is boosted by synthesized data generated using data augmentation, especially copy and paste method.
The qualitative and quantitative results can be found in the presentation and report in /materials/ folder
Install dependencies using anaconda.
conda env create -f environment.ymlThe GPU unit is required.
To overview the functionality of the project, see Jupyter Notebooks. The project consists of three main modules:
- Data Augmentation to synthesize new data
- Image Classification to solve the problem with classification approach
- Image Segmentation to solve the problem with instance segmentation approach.
Before start, note that
- The execution GPU can be defined in image_segmentation/train_net.py and image_classification/transfer_learning.py
- For convinience it is recommended to define store and project directory in image_classification/constants.py