Classifying Plastic Waste on River Surfaces utilising CNN and Tensorflow

Waste in rivers is an ever-increasing problem. This paper will look at Deep Learning and Computer Vision technologies to determine if they can be applied to the problem domain. Usage of Deep Learning and Computer Vision technologies has grown massively in the last few years thanks to increased computational power, the availability of training data such as ImageNet, and the availability more complex and efficient algorithms. This research investigates two models to determine which one is more suited for the problem domain by evaluating their results based on performing training and testing on a developed waste dataset for the purposes of this research. The dataset is developed four times, each variant incurring more implementation of pre-processing techniques than the other. This resulted in the same dataset being tested four times on both models with varying levels of pre-processing. The first variant of the dataset had no pre-processing, the second with aspect ratio adjusting, the third dataset being augmented by the image data generator, and the fourth by way of an independent augmentation pipeline. The developed waste dataset has images of size 100x100 dimensions regardless of variant. Variant one of the waste datasets contained 1000 images and expanded all the way up to 19,973 images after pipeline augmentation in variant 4. Both VGG-16 and DenseNet-201 will have all four variants implemented on them to investigate which CNN best suits this research domain but also investigate the differences of applying different pre-processing techniques and how this affects results yielded by the two CNN models.