High-Level Project Summary
The team has developed a model for detecting common plastic objects in the sea that will traverse the 5 oceanic debris patches found in the "Garbage Patch Visualization Experiment". 487 internet images of bottles, bags and caps have been collected. Data Augmentation techniques (filters, rotation and image size transformation) were applied in order to increase their number. Using Google Colab's python programming environment to configure YOLOV3 and V4 (You Only Look Once), and the pre-trained models darknet53.conv.74 and yolov4.conv.137, a network was developed that is capable of detecting and quantifying plastic marine debris. The model has remarkable precision, which is why it is important.
Link to Project "Demo"
Link to Final Project
Detailed Project Description
Images of plastic waste selected from the internet were selected: dataset TACO (Proença, 2021) and Trashnet (Thung, 2017). A total of 500 images were taken and contain the following categories: bottle, plastic cap, plastic bag, ropes. Additionally, it is worth mentioning that this sample, when applied in a Deep Learning model, will later be partitioned in a train (80%), test (20%) as developed (Kylili, Kyriakides, Artusi, & Hadjistassou, 2019) in his project.
The development of the detection model that will be presented in this research is made up of a series of steps that will be seen in the following diagram.
Stages:
Acquisition
At this stage, it will be necessary to develop a database of images of plastic waste that fall into the following categories: bottle, plastic cap, plastic bag and ropes; all of these will be chosen using the following datasets: TACO (Proença, 2021) and Trashnet (Thung, 2017).
Preprocessing
For this phase, we will try to follow a series of steps to highlight the characteristics of the images, it is important to use the greatest number of Computer Vision techniques: data augmentation, the steps that we would like to implement would be: image preparation, color enhancement, sharpening and resizing of the images, with this aim to have a dataset ready for a phase of retraining and obtaining results.
Modeling and Classification
In this phase we will proceed to create new categories to proceed with re-training in each model to be evaluated, then we will proceed to train the YoloV3 and YoloV4 models, the environment selected to carry out this activity will be that of Google Collaboratory, as it offers us great advantages with respect to time that will be used for this activity, the environment will be configured for use with GPUs.
choose the best detection model, the following indicators were used:
- Accuracy: (TP + TN) / (TP + FP + TN + FN)
- Precision: TP / (TP + FN)
- Recall: TP / (TP + FP)
- F1 Score: (Precision * Recall) / (Precision + Recall)
Resultados:
The following graphs show the evolution of training with respect to the mAP and Loss of each model.
Space Agency Data
-Garbage Patch Visualization Experiment, this study provided us with the 5 key points where the idea of the presented project would be developed.
-"Marine Debris: Impacts on Ecosystems and Species" and "Marine Debris and Wildlife: Impacts, Sources, and Solutions" these articles gave us the idea of the problem.
Hackathon Journey
A good opportunity to test our knowledge.
References
-Kylili, K., Kyriakides, I., Artusi, A., & Hadjistassou, C. (2019). Identifying floating plastic marine debris using a deep learning approach. doi:10.1007/s11356-019-05148-4
-Proença, P. F. (26 de Marzo de 2021). TACO. https://github.com/pedropro/TACO
-Thung, G. (9 de Abril de 2017). Trashnet. Obtenido de Github: https://github.com/garythung/trashnet
-https://www.doi.gov/ocl/marine-debris-impacts#main-content
-https://www.doi.gov/ocl/marine-debris
-https://sos.noaa.gov/catalog/datasets/marine-debris-garbage-patch-experiment-drifters-and-model/
Tags
#water, #computer_Vision, #Deep_Learning, #Convolutional_Neural_Networks.
Global Judging
This project has been submitted for consideration during the Judging process.