Machine learning modeling for image segmentation in manufacturing and agriculture applications
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
This dissertation focuses on applying machine learning (ML) modelling for image segmentation tasks of various applications such as additive manufacturing monitoring, agricultural soil cover classification, and laser scribing quality control. The proposed framework uses various ML models such as gradient boosting classifier and deep convolutional neural network to improve and automate image segmentation tasks. In recent years, supervised ML methods have been widely adopted for imaging processing applications in various industries. The presence of cameras installed in production processes has generated a vast amount of image data that can potentially be used for process monitoring. Specifically, deep supervised machine learning models have been successfully implemented to build automatic tools for filtering and classifying useful information for process monitoring. However, successful implementations of deep supervised learning algorithms depend on several factors such as distribution and size of training data, selected ML models, and consistency in the target domain distribution that may change based on different environmental conditions over time. The proposed framework takes advantage of general-purposed, trained supervised learning models and applies them for process monitoring applications related to manufacturing and agriculture. In Chapter 2, a layer-wise framework is proposed to monitor the quality of 3D printing parts based on top-view images. The proposed statistical process monitoring method starts with self-start control charts that require only two successful initial prints. Unsupervised machine learning methods can be used for problems in which high accuracy is not required, but statistical process monitoring usually demands high classification accuracies to avoid Type I and II errors. Answering the challenges of image processing using unsupervised methods due to lighting, a supervised Gradient Boosting Classifier (GBC) with 93 percent accuracy is adopted to classify each printed layer from the printing bed. Despite the power of GBC or other decision-tree-based ML models comparable to unsupervised ML models, their capability is limited in terms of accuracy and running time for complex classification problems such as soil cover classification. In Chapter 3, a deep convolutional neural network (DCNN) for semantic segmentation is trained to quantify and monitor soil coverage in agricultural fields. The trained model is capable of accurately quantifying green canopy cover, counting plants, and classifying stubble. Due to the wide variety of scenarios in a real agricultural field, 3942 high-resolution images were collected and labeled for training and test data set. The difficulty and hardship of collecting, cleaning, and labeling the mentioned dataset was the motivation to find a better approach to alleviate data-wrangling burden for any ML model training. One of the most influential factors is the need for a high volume of labeled data from an exact problem domain in terms of feature space and distributions of data of all classes. Image data preparation for deep learning model training is expensive in terms of the time for labelling due to tedious manual processing. Multiple human labelers can work simultaneously but inconsistent labeling will generate a training data set that often compromises model performance. In addition, training a ML model for a complication problem from scratch will also demand vast computational power. One of the potential approaches for alleviating data wrangling challenges is transfer learning (TL). In Chapter 4, a TL approach was adopted for monitoring three laser scribing characteristics – scribe width, straightness, and debris to answer these challenges. The proposed transfer deep convolutional neural network (TDCNN) model can reduce timely and costly processing of data preparation. The proposed framework leverages a deep learning model already trained for a similar problem and only uses 21 images generated gleaned from the problem domain. The proposed TDCNN overcame the data challenge by leveraging the DCNN model called VGG16 already trained for basic geometric features using more than two million pictures. Appropriate image processing techniques were provided to measure scribe width and line straightness as well as total scribe and debris area using classified images with 96 percent accuracy. In addition to the fact that the TDCNN is functioning with less trainable parameters (i.e., 5 million versus 15 million for VGG16), increasing training size to 154 did not provide significant improvement in accuracy that shows the TDCNN does not need high volume of data to be successful. Finally, chapter 5 summarizes the proposed work and lays out the topics for future research.