This project's main aim is to determine the level of edge-based processing required to reliably detect the presence and germination of fungal spores a

点击翻译按钮获取中文摘要

This project's main aim is to determine the level of edge-based processing required to reliably detect the presence and germination of fungal spores and determine the share of this processing that can be done on the cloud so that the system is robust and reliable overall. To do this, first the spores will be imaged using a low-cost camera connected to an embedded computer (Raspberry Pi 4), the pixel size of the camera is 1.12 micrometres which should be theoretically sufficient to clearly image both the spores (30 micrometres) and their germ tubes (5-6 micrometres). Another question posed here is, is the camera system adequate to image the desired objects clearly enough to allow for reliable processing further down the pipeline and what level of image processing is required to produce an image of sufficient quality. The approach that will be taken to determine an answer for this will be to first acquire a dataset of images containing the desired objects to be detected (e.g., in this case it is the spores themselves as well as germinated spores with visible germ tubes). This will require a camera rig to be constructed in the lab and spore germination trials to be run in controlled conditions to allow us to ascertain the required conditions for spore germination under the camera. Once the germination rates are reliable then the dataset can be built over many germination trials and techniques such as data augmentation implemented to ensure that there is a high level of variation in the dataset. This will prevent issues such as overfitting if machine learning techniques are used for the final detection/classification task.The next task after this is to use the dataset to train a model to detect the spores and their germ tubes, there are multiple ways this could be done, however two approaches will most likely be evaluated for performance, these being traditional machine learning/deep learning and a technique such as a support vector machine coupled with a clustering technique such as partial least squares regression. These approaches are vastly different and as a result there should be a noticeable difference in performance at the edge, providing a very useful comparison and an insight into the performance and processing cost of each approach. This will allow for the most suitable option to be selected to be deployed onto a low power device for testing. The stage after this will mostly be about optimization and addressing the question of the split in processing between the edge device and the cloud.