Currently, it is becoming more and more difficult to obtain a labor force for the agricultural sector. In combination with machine vision progress and low-cost cameras, this leads to an increase in interest in automation. This automation could be considered in different ways; from a yield prediction based on satellite data to a fully autonomous robot driving around on the field. Nevertheless, for some of these automated implementations, machine vision using cameras is required. This machine vision needs to reliably represent the field of view to be applied for automation. Therefore, an estimation of the quality of the field of view is required; for example detecting whether there is dirt on the camera, and yes, quantifying the influence on the quality of the camera view.
The task of this project is to investigate how the field of view quality of a camera can be assessed and subsequently quantified. For this project, several algorithms could be compared, such as the laplacian value or a deep learning algorithm. Subsequently, a dataset needs to be collected, which includes common aspects negatively influencing the vision quality. For the dataset collection, you could think about putting a small amount of dirt on your camera and trying to detect and quantify this dirt. Furthermore, the dataset should consist of images collected from different datasets, containing several images which should not be classified as images with low machine vision quality. At the end of the project, you should have a quantified analysis of the different algorithms and describe the advantages and disadvantages per algorithm.
20% Theory, 10% data collection, 50% implementation, 20% writing report
Le Viet Duc, email@example.com