Thesis Format
Integrated Article
Degree
Master of Engineering Science
Program
Electrical and Computer Engineering
Supervisor
Eagleson, Roy
2nd Supervisor
De Ribaupierre, Sandrine
Co-Supervisor
Abstract
This manuscript-style thesis covers a series of experiments that incorporate object detection and seeks new insights. The first project shines a light on the relatively obscure field of temporal image ordering and seeks to understand whether injecting tensors with explicit summaries of object detection model predictions could improve the performance of Siamese Networks tasked with temporal ordering. The expectations of improved results after the addition of summary tensors generated by object detection models to the convolutional embeddings were supported by the results of the experimentation. The second project involves a new spin on an old theory, Anne Triesman laid out in her Feature Integration Theory a set of relationships that described some aspects of human visual search in conjunction and feature search problems. This exploratory and comparative study seeks to find analogous performance profiles in object detection models as shown by humans when completing the same types of problems. Our findings underscore the importance of validation in drawing meaningful insights from comparative studies between humans and AI models. The methodology proves compelling as a future avenue of further comparison between artificial intelligence models and humans, however, this particular set of models did not provide an analogous performance profile. The third project is a medically sensitive project which requires ensuring data privacy. Endoscopic third ventriculostomy (ETV) is a critical neurosurgical procedure used in managing hydrocephalus. This paper attempts to generate and compare artificial intelligence-based tools that seek to replicate the experience of an experienced neurosurgeon in determining the optimal location for fenestration during ETV. Leveraging YOLO and RT-DETR object detection models, we compare their respective performances with one another. The findings underscore the promising role of AI-based tools in augmenting surgical practices and enhancing patient outcomes. The fourth project outlines a learning exercise for myself that takes no consideration towards publication. I built a DEtection TRansofrmer (DETR) using a variety of online resources and bench marked it on the MS COCO dataset. This exercise was designed for me to expand my horizons and allowed me to learn how transformers complete visual processing and generate object detection predictions.
Summary for Lay Audience
The field of object detection has experienced large leaps in edge compatibility, speed and accuracy in recent years. In following these developments its important to consider unique use cases of these new object detection models in a variety of interesting use cases. This manuscript style thesis covers a series of experiments that incorporate object detection and seeks new and interesting insights. Including; designing a tool to aid surgeons, making a custom object detection model using a transformer, contrasting visual search performance profiles between humans and object detection models and finally predicting which image happened first while exploring a three modalities (two of them relying on object detection).
Recommended Citation
Chandel, Chandan, "Machine Learning Methods and Comparisons for Visual Search, Ordering, Detection and Localization with Applications for Medical Imaging" (2024). Electronic Thesis and Dissertation Repository. 10435.
https://ir.lib.uwo.ca/etd/10435
Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License