Date of Graduation

12-2024

Document Type

Thesis

Degree Name

Master of Science in Food Science (MS)

Degree Level

Graduate

Department

Food Science

Advisor/Mentor

Wang, Dongyi

Committee Member

Subbiah, Jeyamkondan

Second Committee Member

Owens, Casey M.

Third Committee Member

Koparan, Cengiz

Keywords

Bayesian Optimization (BO); Hardness Distribution Map; Hyperparameter tuning; Neural Network Architecture Search; wide and deep learning; woody breast

Abstract

The development and implementation of a Wide & Deep (WD) learning model tailored for classification and regression tasks utilizing spectral data provides a robust solution to evaluate woody breast (WB) conditions in poultry fillets. This process begins with thorough data preprocessing, which includes loading spectral and classification datasets, imputing missing values with medians, and splitting the data into training and testing sets to ensure rigorous model evaluation. The WD model architecture integrates wide linear models and deep neural networks to harness the strengths of both approaches. The wide component excels at memorizing sparse feature interactions, while the deep component captures generalized and non-linear feature interactions through multiple layers. This design ensures the model's ability to handle both known and unseen feature relationships effectively.
To enhance model performance, the implementation leverages TensorFlow and Keras libraries alongside Keras Tuner for hyperparameter optimization. The hyperparameter tuning is guided by Bayesian optimization, enabling the systematic selection of optimal configurations. Training is monitored with an early stopping mechanism, halting the process when validation performance plateaus, thereby preventing overfitting and maintaining generalizability. The final model is evaluated using accuracy metrics, confusion matrices, and regression correlation, providing a comprehensive assessment of its classification and regression capabilities.
The WB condition, which has emerged as a significant challenge in the poultry industry due to intensive genetic selection for rapid growth and high broiler yields, has led to economic losses exceeding $200 million annually. While human palpation remains the most common method for WB detection, it is labor-intensive, subjective, and inconsistent. Hyperspectral imaging (HSI), coupled with machine learning, offers a non-invasive, objective, and high-throughput alternative for evaluating WB conditions. In this study, 106 raw fillet samples (categorized as normal, mild, and severe WB) were analyzed using spectral data, with the spatially heterogeneous distribution of hardness explicitly considered in the model design.
The novel Wide & Deep model integrates neural architecture search (NAS), named NAS-WD, to automate the optimization of network architecture and hyperparameters. This model not only classified WB levels with an overall accuracy of 95.00%, outperforming traditional machine learning models, but also established a regression model correlating spectral data with sample hardness, achieving a correlation coefficient (r) of 0.75. These results highlight NAS-WD’s superior ability to address classification and regression tasks compared to conventional approaches, offering an innovative solution to an industry-wide problem.

Download

Share

COinS