Deep learning for industrial applications
<p>Improvement of industrial quality control processes is paramount to support the ever-growing world population. A key part of this process has traditionally relied on classical vision techniques. With the recent success of deep learning in many areas of vision, it is envisioned that its appl...
| Prif Awdur: | |
|---|---|
| Awduron Eraill: | |
| Fformat: | Traethawd Ymchwil |
| Iaith: | English |
| Cyhoeddwyd: |
2023
|
| _version_ | 1826310805137653760 |
|---|---|
| author | Data, GWP |
| author2 | Prisacariu, V |
| author_facet | Prisacariu, V Data, GWP |
| author_sort | Data, GWP |
| collection | OXFORD |
| description | <p>Improvement of industrial quality control processes is paramount to support the ever-growing world population. A key part of this process has traditionally relied on classical vision techniques. With the recent success of deep learning in many areas of vision, it is envisioned that its application to this process may bring about significant improvements. The aim of this thesis is to develop and validate how a deep learning based solution may be incorporated into this process.</p>
<p>To this end, we considered how a deep learning solution can be integrated into a widely-used process architecture, and specified a system design that we implemented and validated in the lab.</p>
<p>We then identified data labelling to be a potential bottleneck in extending the solution to other applications, which led us to develop few-shot methods to address this. First, we introduced a method for interpolating batch normalization parameters across class-specialized networks for few-shot image classification. This significantly reduces overfitting compared to fine-tuning and allows for efficient adaptation, which we benchmark against CIFAR10 and ImageNet32. Second, we proposed a cosine-similarity-based method for online few-shot detection. Cosine similarity is used to produce detection scores of unseen exemplars against test candidates without retraining the network. We demonstrate this on the few-shot variant of ImageNet and VOC.</p> |
| first_indexed | 2024-03-07T07:57:25Z |
| format | Thesis |
| id | oxford-uuid:5d0f002c-d39b-441d-9e1d-321380fcb338 |
| institution | University of Oxford |
| language | English |
| last_indexed | 2024-03-07T07:57:25Z |
| publishDate | 2023 |
| record_format | dspace |
| spelling | oxford-uuid:5d0f002c-d39b-441d-9e1d-321380fcb3382023-09-05T14:33:39ZDeep learning for industrial applicationsThesishttp://purl.org/coar/resource_type/c_db06uuid:5d0f002c-d39b-441d-9e1d-321380fcb338EnglishHyrax Deposit2023Data, GWPPrisacariu, V<p>Improvement of industrial quality control processes is paramount to support the ever-growing world population. A key part of this process has traditionally relied on classical vision techniques. With the recent success of deep learning in many areas of vision, it is envisioned that its application to this process may bring about significant improvements. The aim of this thesis is to develop and validate how a deep learning based solution may be incorporated into this process.</p> <p>To this end, we considered how a deep learning solution can be integrated into a widely-used process architecture, and specified a system design that we implemented and validated in the lab.</p> <p>We then identified data labelling to be a potential bottleneck in extending the solution to other applications, which led us to develop few-shot methods to address this. First, we introduced a method for interpolating batch normalization parameters across class-specialized networks for few-shot image classification. This significantly reduces overfitting compared to fine-tuning and allows for efficient adaptation, which we benchmark against CIFAR10 and ImageNet32. Second, we proposed a cosine-similarity-based method for online few-shot detection. Cosine similarity is used to produce detection scores of unseen exemplars against test candidates without retraining the network. We demonstrate this on the few-shot variant of ImageNet and VOC.</p> |
| spellingShingle | Data, GWP Deep learning for industrial applications |
| title | Deep learning for industrial applications |
| title_full | Deep learning for industrial applications |
| title_fullStr | Deep learning for industrial applications |
| title_full_unstemmed | Deep learning for industrial applications |
| title_short | Deep learning for industrial applications |
| title_sort | deep learning for industrial applications |
| work_keys_str_mv | AT datagwp deeplearningforindustrialapplications |