This work is dedicated to the development and improvement of the neural networks in a context of semantic segmentation, based on U-Net architecture, which allows for improved overall evaluation and performance metrics, compared to the baseline U-Net model.
The topic of the dissertation was agreed upon by the Department of Computer Engineering of the National Technical University of Ukraine “Ihor Sikorskiy Kyiv Polytechnic Institute”, in accordance with Cabinet of Ministers Order №1556-p on the concept of development of Artificial Intelligence in Ukraine.
The goal of the dissertation was an improvement of existing semantic image segmentation methods, that allow to improve results and effectiveness of neural networks.
To achieve the stated goal, the next tasks and problems were solved:
- Existing semantic image segmentation methods and neural network architectures.
- The family of U-Net neural network architectures was studied in detail.
- As part of the proposed method, expansion rate variation and depthwise separable skip connections were proposed as modifications to the baseline U-Net architecture.
- Many experiments were conducted on different datasets, using different approaches and proposed novel improvements. To confirm the quality nature of the change, K-fold cross-validation was performed;
- The benchmark was conducted for the proposed improvements and modifications to get inference time and memory usage metrics.
The proposed expansion rate variation method allows the developer to regulate the number of parameters for the U-Net network, enabling deeper networks with fewer parameters. Using this approach, it is possible to optimize the size and the performance of the model and get the same evaluation results as a baseline model, with 2.5 fewer parameters.
Another proposal is using Depthwise Separable Skip Connections, based on Depthwise Separable Convolutions. This modification allowed to improve the results of the model with a relatively small size increase for a model. It is also possible to improve other U-Net-like models, as demonstrated in the Attention-UNet model. For different datasets, at least one of the suggested modifications managed to improve a result of the baseline model, with this improvement measured between 1-5%. It was also shown on the K-fold cross-validation, that these modifications could steadily outperform the baseline model. In some cases, this improvement was achieved by a model with just a 1% increase in a number of parameters, which proves this is a quality improvement.
Utilizing the proposed modifications, a method for image segmentation using modified U-Net architectures was developed with Python programming language and TensorFlow library, to prove the feasibility of these modifications. Experiments were conducted in different knowledge domains, such as Computer Assisted Diagnostics and road environment analysis. Also, these methods were used for analyzing both 2-dimensional images and 3-dimensional volumes, which proves the practicality of using these methods. For the experiments, well-known state-of-the-art datasets were used, such as the University of Wisconsin Gastrointestinal (UWGIT), Cityscapes, Synapse, and Brain Tumor Segmentation (BraTS). It was also demonstrated that the proposed methods may match, and sometimes outperform some of the State-Of-The-Art models, aside from baseline U-Net.
Besides that, some performance and productivity metrics were measured for the proposed network architectures. It was noted, that deeper networks with an expansion rate approach may have better inference times than a baseline model and match its quality.
The suggested methods have a big practical value and a wide range of applications in the field of image analysis, which was proven during the conducted experiments.
