Traditional Image Processing and Modern Computer Vision Techniques for the Study of Two-Phase CO2 Flow

Master Thesis


Permanent link to this Item
Journal Title
Link to Journal
Journal ISSN
Volume Title
The work presented here details the development of software-based tools for the extraction of physical parameters which describe two-phase (gas-liquid) upwardly flowing CO2, for the purpose of using these parameters as sensor data for a control feedback loop, and for the automatic detection of flow regime transition, which is useful for the development of flow regime maps. The core focus of this thesis is the development of these tools in such a way that their primary input is an image or set of images. To achieve this, two schools of thought are explored: First, traditional image processing techniques are employed to study the flow. These techniques require manual image feature selection, and they make use of purposebuilt algorithms to extract the desired parameters from an input image using these features. The second approach makes use of modern computer vision techniques, where the image features are automatically learnt through machine learning, and an end-to-end network design makes use of these features to extract the desired output without manual tuning. Traditional image processing is used to develop an algorithm which extracts the void fraction value from an image of bubbly flow. This algorithm works by detecting individual bubbles within the input image, and then estimating the volume of each bubble (with uncertainty) in order to calculate the final void fraction. The outputs seen from this algorithm correlated well with those produced by established models for calculating void fraction, but the problem with this algorithm is its limited scope of use: it is only applicable to images of bubbly flow, a flow regime which exists for only a small portion of the total possible vapour quality range under steady state conditions. Two different tools, which share a similar architecture, and which classify flow regime and vapour quality respectively, were successfully developed using modern computer vision techniques. These models both take in video clips as their inputs. The approach makes use of deep learning to train a convolutional neural network (CNN), which is used for individual frame classification and image feature extraction, and a deep long short-term memory (LSTM) network, used to capture temporal features present in a sequence of image feature sets, and to make a final vapour quality or flow regime classification. The proposed architecture achieves accurate flow regime and vapour quality classifications in practical application to two-phase CO2 flow in vertical tubes based on off-line data and an on-line prototype implementation, developed as a proof of concept for the use of these models within a feedback control loop. The successful application of the LSTM network reveals the significance of temporal information for image based studies of multi-phase flow. When comparing these parallel developments, the advantages and disadvantages of the two approaches can be clearly seen. Traditional image processing requires far more extensive domain specific knowledge and manual fine tuning, but this approach allows for a user to clearly understand the outputs of the algorithm, whether they are correct or incorrect, as the internal mechanisms of the algorithm are all purpose built. This is not the case for deep learning based modern computer vision methodologies, which are more of a “black box”. These methods require a large amount of training data, but less domain specific knowledge, as the important features from the input data do not need to be manually selected and processed by the user. This leads to high performing systems which are difficult to understand and debug. There are different cases in which each of these methods would be preferable, but with the rapid evolution of deep learning and computer vision over the last few years, deep learning based computer vision appears to be replacing the traditional approach in many cases.