Hyperspectral Imaging and Deep Learning for Food Safety Assessment
| aut.embargo | No | en_NZ |
| aut.thirdpc.contains | No | en_NZ |
| dc.contributor.advisor | Klette, Reinhard | |
| dc.contributor.advisor | Yan, Wei Qi | |
| dc.contributor.author | Al-Sarayreh, Mahmoud | |
| dc.date.accessioned | 2020-07-09T01:28:36Z | |
| dc.date.available | 2020-07-09T01:28:36Z | |
| dc.date.copyright | 2020 | |
| dc.date.issued | 2020 | |
| dc.date.updated | 2020-07-09T00:35:39Z | |
| dc.description.abstract | Hyperspectral imaging (HSI) systems are valuable tools for merging both spectroscopic and computer vision technologies in a single system with the advantage of providing the physical attributes and chemical distribution of materials represented in an image. Snapshot HSI systems are portable systems enabling the generation of HSI images at the video rate with limited spectral information, which makes this technology a step toward real-time HSI applications. For food processing, HSI systems are considered to form a rapid, non-invasive, non-destructive and chemical-free technology for predicting food attributes regarding food safety and quality assessment, which reflects positively on costs, accuracy and processing time of applications in the food industry. This thesis studies the interaction between chemical and textural distributions, presented as spectral and spatial features of HSI images. Moreover, the thesis discusses several traditional and novel approaches in computer vision and deep learning for utilising this interaction in applications of safety assessment of food such as adulteration detection in meat products, authenticity of meat products and foreign object detection (FOD) in meat products. In application of adulteration detection, traditional approaches are investigated for detecting the adulteration in meat products including spectral features and handcrafted textural features obtained from HSI images. Moreover, this thesis presents a novel multi-structure deep learning model for self-feature extraction and combination of these distributions (i.e., chemical and textural) in a single prediction model by using convolution neural networks (CNN). The model is evaluated against the traditional approaches and showed efficiency in prediction. Extraction of joint spectral and spatial features from HSI images is also discussed in this thesis. A 3D–CNN approach is proposed for extracting the joint features. Red-meat classification (case study of fine-grain material classification) is used for evaluating the proposed approach. Moreover, we propose a novel graph-based postprocessing method for enhancing the prediction of the 3D–CNN approach or of any pixel-wise classification model. The proposed classification framework is evaluated against traditional machine learning algorithms such as support vector machines and partial least square discriminant analysis. Three datasets were collected for the evaluation by using three HSI systems: line scanning, near-infrared (NIR) snapshot and visible (VIS) snapshot HSI. In application of FOD in meat products, the thesis discusses the object detection problem in HSI images based on their spectral and spatial features. A novel sequential deep learning framework is proposed for foreign object localization and classification by using CNN networks. The framework includes three modules in a sequential flow: Region proposal, filtering and classification modules. Two independent datasets of NIR snapshot HSI images, contaminated by many types of foreign materials, were used for training and testing the proposed approach. The evaluation showed promising efficiency of the proposed framework in terms of accuracy and real-time processing, compared with a baseline method for FOD such as the selective search approach. | en_NZ |
| dc.identifier.uri | https://hdl.handle.net/10292/13514 | |
| dc.language.iso | en | en_NZ |
| dc.publisher | Auckland University of Technology | |
| dc.rights.accessrights | OpenAccess | |
| dc.subject | Food processing | en_NZ |
| dc.subject | Meat processing | en_NZ |
| dc.subject | Adulteration detection | en_NZ |
| dc.subject | Meat authenticity | en_NZ |
| dc.subject | Hyperspectral imaging | en_NZ |
| dc.subject | HSI | en_NZ |
| dc.subject | Snapshot HSI | en_NZ |
| dc.subject | Material classification | en_NZ |
| dc.subject | Spectral | en_NZ |
| dc.subject | Spatial | en_NZ |
| dc.subject | Textural | en_NZ |
| dc.subject | Joint features | en_NZ |
| dc.subject | Object detection | en_NZ |
| dc.subject | Foreign object | en_NZ |
| dc.subject | FOD | en_NZ |
| dc.subject | Foreign bodies | en_NZ |
| dc.subject | Deep learning | en_NZ |
| dc.subject | 2D-CNN | en_NZ |
| dc.subject | 3D--CNN | en_NZ |
| dc.subject | Region proposal | en_NZ |
| dc.subject | RPN | en_NZ |
| dc.subject | Classification | en_NZ |
| dc.title | Hyperspectral Imaging and Deep Learning for Food Safety Assessment | en_NZ |
| dc.type | Thesis | en_NZ |
| thesis.degree.grantor | Auckland University of Technology | |
| thesis.degree.level | Doctoral Theses | |
| thesis.degree.name | Doctor of Philosophy | en_NZ |