Year: 2021
Type: dissertation
Abstract: U-Net, which was suggested for precise pixel-wise segmentation of biomedical images, is one of the excellent image-to-image transformation models.Due to its effective pixel-wise localization representation capability, U-Net is widely used in various research fields.There are lots of studies for improving the performance of U-Net based models or training U-Net based models with insufficient data.In this paper, two novel methods, which are based on U-Net, are introduced.First, cascading modular U-Nets (CMU-Nets) are suggested to train the model by using a limited number of training data.CMU-Nets are organized with modularized U-Nets trained by a sufficient amount of different training data similar to the target data.The modularized pre-trained U-Nets are arranged by the optimal order and cascaded as an end-to-end system.The cascaded modularized U-Nets are connected by the proposed novel intermodule skip-connections for transferring pixel-wise localization information between the modules.Consequently, CMU-Nets are trained by the target data as a fine-tuning process.Second, for improving the performance of CMU-Nets by attention mechanism, an effective improvement method that adopts Squeeze and Excitation network (SE-Net) is introduced in CMU-Nets.The performance of conventional CNN-based models can be improved by unconstrained SE-Net utilization.However, CMU-Nets comprise pre-trained U-Net modules and are connected by the inter-module skipconnections, a novel method of SE-Net utilization should be suggested.Therefore, the proposed method, which can be called Attention CMU-Nets (ACMU-Nets), is suggested by reflecting these characteristics of CMU-Nets.To demonstrate the performance of these two proposed methods, diverse quantitative and qualitative analysis results are introduced.Since these two proposed methods are designed to address complex image transformation issues, the methods are trained by using the Document Image Binarization Competition (DIBCO) and Handwritten DIBCO (H-DIBCO) dataset for complex historical document image binarization.
Author: 姜, 碩埈
Institution: Kyushu University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7155664896
