Master of Science
Electrical & Computer Engineering
The objective of computer vision research is to endow computers with human-like perception to enable the capability to detect their surroundings, interpret the data they sense, take appropriate actions, and learn from their experiences to improve future performance. The area has progressed from using traditional pattern recognition and image processing technologies to advanced techniques in image understanding such as model-based and knowledge-based vision. In the past few years there has been a surge of interest in machine learning algorithms for computer vision-based applications. Machine learning technology has the potential to significantly contribute to the development of flexible and robust vision algorithms that will improve the performance of practical vision systems with a higher level of competence and greater generality. Additionally, the development of machine learning-based architectures has the potential to reduce system development time while simultaneously achieving the above-stated performance improvements. This work proposes the utilization of a computer vision-based approach that leverages machine and deep learning systems to aid the detection and identification of sow reproduction cycles by segmentation and object detection techniques. A lightweight machine learning system is proposed for object detection to address dataset collection issues in one of the most crucial and potentially lucrative farming applications. This technique was designed to detect the vulvae region in pre-estrous sows using a single thermal image. In the first experiment, the support vector machine (SVM) classifier was used after extracting features determined by 12 Gabor filters. The features are then concatenated with the features obtained from the Histogram of oriented gradients (HOG) to produce the results of the first experiment. In the second experiment, the number of distinct Gabor filters used was increased from 12 to 96. The system is trained on cropped image windows and uses the Gaussian pyramid technique to look for the vulva in the input image. The resulting process is shown to be lightweight, simple, and robust when applied to and evaluated on a large number of images. The results from extensive qualitative and quantitative testing experiments are included. The experimental results include false detection, missing detection and favorable detection rates. The results indicate state-of-the-art accuracy. Additionally, the project was expanded by utilizing the "You Only Look Once" (YOLO) deep learning Object Detection models for fast object detection. The results from object detection have been used to label images for segmentation. The bounding box from the detected area was systematically colored to achieve the segmented and labeled images. Then these segmented images are used as custom data to train U-Net segmentation. The first step involves building a machine learning model using Gabor filters and HOG for feature extraction and SVM for classification. The results discovered the deficiency of the model, therefore a second stage was suggested in which the dataset was trained using YOLOv3-dependent deep learning object detection. The resulting segmentation model is found to be the best choice to aid the process of vulva localization. Since the model depends on the original gray-scale image and the mask of the region of interest (ROI), a custom dataset containing these features was obtained, augmented, and used to train a U-Net segmentation model. The results of the final approach shows that the proposed system can segment sow's vulva region even in low rank images and has an excellent performance efficiency. Furthermore, the resulting algorithm can be used to improve the automation of estrous detection by providing reliable ROI identification and segmentation and enabling beneficial temporal change detection and tracking in future efforts.
Dissertation or thesis originally submitted to ProQuest
Almadani, Iyad, "ORGAN LOCALIZATION AND DETECTION IN SOW’S USING MACHINE LEARNING AND DEEP LEARNING IN COMPUTER VISION" (2023). Electronic Theses and Dissertations. 3104.