Electronic Theses and Dissertations

Iyad Almadani

Date

2025

Document Type

Dissertation

Degree Name

Doctor of Philosophy

Department

Electrical & Computer Engineering

Committee Chair

Aaron Robinson

Committee Member

Deepak Venugopal

Committee Member

Eddie Jacobs

Committee Member

Madhusudhanan Balasubramanian

Abstract

Traditional methods for predicting sow reproductive cycles are expensive, labor-intensive, and expose workers to occupational hazards, including respiratory toxins, repetitive stress injuries, and mental health issues. On large farms with limited staff, managing individual sow health becomes increasingly difficult, underscoring the need for automated solutions. This is where computer vision offers a promising, non-contact method to detect estrus by analyzing vulva features like size and using thermal imagery to monitor temperature changes, which are key indicators of reproductive readiness. Automating this process enhances breeding efficiency by ensuring insemination occurs at the optimal time while continuously monitoring sows and alerting staff to anomalies for early intervention. A critical aspect of automating estrus detection is the accurate analysis of vulva size and temperature variations, as these subtle changes signal ovulation. Maintaining consistent camera distance during data collection is essential for achieving reliable results, as variations in distance can distort measurements, leading to inaccurate estimations of estrus and resulting in missed breeding opportunities or false positives. This inconsistency is particularly problematic in thermal image analysis, where differences in resolution can cause pixel intensity values to misrepresent actual temperature variations. To improve the accuracy and generalizability of computer vision models, researchers must address these challenges by standardizing image capture protocols and calibrating equipment to account for external factors such as atmospheric conditions. Overcoming these limitations will ensure that automated estrus detection remains a reliable and scalable solution that improves sow reproductive health, optimizes farm productivity, and safeguards worker well-being. This dissertation details the comprehensive process of automating estrus detection in sows, starting with the accurate detection of vulva shape and size, which is essential for reliable estrus identification. To achieve this, various methods were proposed to determine the most effective approach for vulva segmentation. The initial chapter explores multiple object segmentation techniques, beginning with the manual annotation of thermal images to enhance the visibility of the Region of Interest (ROI) and ensure precise segmentation. To optimize vulva segmentation, the system was trained using a series of models, evolving from U-Net for semantic segmentation to YOLOv8 and ultimately YOLOv9, as part of a continuous effort to enhance performance and accuracy. The second chapter refines the detection framework by integrating YOLO for vulva segmentation and keypoint detection, allowing for geometric assessments of vulva size and shape. Building on the precise vulva segmentation in the previous chapter, we gained critical insights into its shape, enabling accurate perimeter measurements. Keypoint identification facilitated precise Euclidean distance calculations—horizontally between the labia and vertically from the clitoris to the perineum. Camera calibration through monocular depth estimation linked image-based measurements to real-world distances, ensuring accurate scaling and analysis. Finally, we introduce a classification method to distinguish estrus from non-estrus states by comparing pixel-based dimensions and perimeter measurements. This approach applies a nearest-neighbor algorithm, aggregating distance calculations to classify new data points based on similarity to reference datasets, ensuring reliable estrus detection. Building upon the foundation established in the previous chapters, the final chapter enhances the detection model by further refining YOLO-based segmentation and expanding the analysis to incorporate pixel intensity values from thermal images. This additional layer of analysis leverages the subtle thermal variations in the vulva region, providing crucial insights into estrus detection that were not captured through geometric measurements alone. By integrating pixel intensity as a complementary factor, the model achieves greater sensitivity and accuracy in distinguishing between estrus and non-estrus states. This holistic approach enhances the robustness of the detection framework, ensuring more precise and reliable results across varying environmental conditions and improving the overall effectiveness of automated estrus monitoring systems. To validate our approach, we implemented a three-stage evaluation process. In the first stage, we computed the Mean Squared Error (MSE) to compare the ground truth keypoints of labia distance with the predicted keypoints, applying the same method to measure the distance between the clitoris and perineum. Next, we conducted a quantitative analysis to benchmark the performance of our YOLOv9 segmentation model against the previous U-Net and YOLOv8 models. Finally, we evaluated the classification process by constructing a confusion matrix and comparing the outcomes across different methods. This comprehensive evaluation highlights the advancements in accuracy and performance achieved through our updated model, underscoring its potential for more reliable and automated estrus detection in sows. Through rigorous experimentation and quantitative evaluation, this dissertation demonstrates the transformative potential of computer vision in livestock management. The proposed system offers an automated, non-contact solution for estrus detection, contributing to more efficient reproductive health monitoring and sustainable agricultural practices.

Comments

Data is provided by the student.

Library Comment

Dissertation or thesis originally submitted to ProQuest.

Notes

Open Access

Download

Share

COinS