@phdthesis{8635328,
  abstract     = {{With the wide use of imaging systems in scientific studies, automated image analysis has attracted increasing attention over the past decades. Automated image analysis can extract meaningful information from raw images and generate novel insight, having superiority in processing efficiency and assessment objectivity over manual image analysis. Since many image analysis tasks can be essentially viewed as problems of feature mapping, feature extraction plays a pivotal role in automated image analysis. Feature extraction methods can be generally divided into hand-crafted methods and deep learning methods. Hand-crafted methods have comparatively better interpretability, and their processing procedures can be well controlled by explicitly defined parameters. Representative hand-crafted features include edges, lines, and blobs. Deep learning methods have a powerful ability to learn and represent features, and thus, they can accomplish comparatively more complex tasks. Hand-crafted methods and deep learning methods are not mutually exclusive. Many deep learning methods have greatly benefited from hand-crafted features. Therefore, it is still necessary to further explore the potential of hand-crafted features to have more explicit toolkits, and to provide more explanations for deep learning techniques. Moreover, it would be worth developing methods that jointly exploit the advantages of hand-crafted methods and deep learning methods.
In bioscience engineering, although a large amount of image analysis methods have been developed, there are still many particular tasks that remain unsolved. In aquaculture, the brine shrimp Artemia is a widely used cost-effective diet for fish and crustaceans, and recently, the number of Artemia studies is increasing. Since Artemia objects are very small in size, they are usually observed by a stereo-microscope, which can acquire a large amount of Artemia images containing many Artemia objects. Conventionally, most of the Artemia image analysis tasks are carried out manually, which is rather time-consuming and labor-intensive. Hence, it is quite necessary to design tailor-made methods for analyzing Artemia images.  
In this thesis, on the one hand, we investigate several Gaussian-based convolutional kernels to extract hand-crafted features. We present the normalized first-order derivative of anisotropic Gaussian kernel that can detect multiscale edges with a good noise-robustness. This kernel is applied to contour detection and superpixel segmentation. Also, we develop a line detection method using the normalized and adaptive second-order anisotropic Gaussian kernel. This method can effectively detect multiscale lines in a noisy environment. For blob characterization, we propose the unilateral second-order Gaussian kernel that can quantitatively measure the blob prominence, scale, and position, while yielding little response for non-blob structures. The favorable properties of this kernel are confirmed in image denoising. Moreover, we propose a blob detection method using iterative Laplacian-of-Gaussian filtering and the unilateral second-order Gaussian kernel. This method can handle overlapping blobs effectively. On the other hand, aiming at automated Artemia image analysis, we propose an Artemia detection and counting method using U-shaped fully convolutional networks and deep convolutional networks. Besides, by jointly using deep learning techniques and hand-crafted features, we develop an automated method that can measure the Artemia length accurately in images. All the proposed methods are validated on either widely adopted datasets or in-house datasets.}},
  author       = {{WANG, Gang}},
  isbn         = {{9789463572514}},
  language     = {{eng}},
  pages        = {{XXXIII, 239}},
  publisher    = {{Ghent University. Faculty of Bioscience Engineering}},
  school       = {{Ghent University}},
  title        = {{Automated image analysis using Gaussian-based convolutional kernels and deep convolutional networks}},
  year         = {{2019}},
}