Application and Comparative Analysis of Different Deep Learning Methods for the Detection of Visual Pollution from Google Street View Images.

Pollution is a pernicious component of the environment. It destroys the ambiance. Because of the pollution, our environment is polluting day by day and ambiances are being uninhabitable. One of the major forms of pollution is visual pollution We have eyes and we are constantly seeing different things with our eyes. And just think whatever our eyes touch should be beautiful, tasteful and appealing. But unfortunately, it is not the actual scenario because we are visually polluting our environment very badly. It is an aesthetic issue which is often overlooked. It is very alarming that we are very irresponsible regarding visual pollution. We should not pollute our environment visually which is trashy and not soothing for our eyes. When people go out for their daily tasks, they can see unwanted pollution all around them. People become very upset and disappointed with the pollution around them because it is not soothing for their eyes. This kind of pollution becomes a curse for people, society, and nation. This study’s objective was to use deep learning to identify visually polluted objects. So, using Google Street View, we compiled our own dataset of 700 images. The dataset was used to train on YOLOv5s, YOLOv5x and YOLOv7 models after being annotated then oversampling techniques were applied to balance the dataset and later on augmented to expand the dataset. The model produced admirable results and had strong predictive power for the selected 6 classes that we had in our dataset. The 6 classes were - Billboards, Street Litters, Construction Materials, Bricks, Wires and Towers. The mAP score for three YOLO models have (Mean Average Precision) score of 86.9%, 88.3% and 89.3% respectively.
TECHNOLOGY::Other technology::Environmental engineering
Department Name
Printed Thesis