Plastic pollution is a significant global environmental challenge, particularly in our oceans.
With an astonishing 5.25 trillion pieces of plastic and microplastics currently floating in the ocean, and approximately 15% of them ending up on our beaches, the impact of plastic pollution cannot be ignored.
To address this pressing issue, our tech team has been actively working on innovative approaches, such as developing a deep learning plastic detection model to map and address plastic pollution on beaches.
Our journey began as part of the 2021 HKGTC iGEM team, where we recognized the urgent need to tackle plastic pollution. To develop effective prevention and cleanup strategies, we collected 718 images of coastal areas, including beaches in Hong Kong such as Cheung Chau and Cheung Sha, using drones and smartphones.
These images were meticulously annotated using CVAT, a computer vision annotation tool provided by Clearbot, with the assistance of dedicated student volunteers.
For plastic detection, we implemented the Mask-RCNN algorithm, a powerful object detection algorithm capable of generating accurate masks and classifying objects. To operationalize our model, we utilized Detectron2 as a framework and leveraged the baselines from Detectron's Model Zoo for transfer learning, resulting in improved model performance.
Additionally, we applied data augmentations to enhance the model's robustness. Through rigorous training, involving 1000 iterations, we achieved promising results, surpassing the overall mean average precision in the Mask-RCNN paper.
Although our initial results demonstrated high average precision, we acknowledged the presence of false positives in the detections.
To address this, we continued to develop our model after the 2021 iGEM. We expanded the plastic category from solely focusing on PET plastic to a wider range of plastics, including plastic buckets, plastic bags, bottle caps, etc.
Furthermore, we increased the dataset size from 718 images to 10,000 images, obtaining the additional images from online sources. We also organized data science experience courses to recruit and train volunteers to assist in annotating the images, further refining the model's accuracy.
Presently, we continue to work diligently to enhance the model's performance. By harnessing the power of deep learning models, we aim to provide valuable data that enables governments, councils, and NGOs to gain comprehensive insights into the current plastic pollution challenges and evaluate the effectiveness of their proposed interventions.
The data provided will empower researchers to develop and implement targeted prevention and cleanup strategies, guiding their efforts towards areas where they can have the greatest impact. Ultimately, our goal is to contribute to the reduction of plastic pollution.
Comments