Miraziz Salehin* at the Department of Computer Science and Engineering, Daffodil International University, has been accepted as satisfactory in partial fulfillment of the requirements for the degree of B.Sc. Department of Computer Science and Engineering Faculty of Science and Information Technology Daffodil International University. We hereby declare that this project was carried out by us under the supervision of Dr.
Sheak Rashed Haider Noori, Associate Professor and Associate Head of Department, CSE Daffodil International University. We also declare that neither this project nor any part of this project has been submitted elsewhere for the award of any degree or diploma. Sheak Rashed Haider Noori, Associate Professor and Associate Head, Department of CSE, Daffodil International University, Dhaka, deserves our deep gratitude and recognition.
Zahid Hasan, Associate Professor, Department of CSE, for his invaluable assistance in completing our project, as well as to the other academic members and employees of Daffodil International University's CSE Department. In such circumstances, we can say that our result is divided into different contexts of the models in order of our result convention. In the case of our result estimation, the best performance was obtained by the DeepLabV3+ ResNet101, the dice loss is 0.05 and the average IoU 90%.
We hope that our work will be enlightened in the case of image segmentation and computer vision massive enrollment.
Introduction
- Motivation
- Rational of the study
- Research Questions
- Expected Output
- Report Layout
The amount of information that can pass through the shortcut is controlled by these parameterized gates. However, tests reveal that Highway Network does not outperform Resent, which is strange because Resent is present in the optimal solution of Highway Network [7], so it should perform at least as well as Resent. In the case of creating or working with deep learning models, scientists have always faced an all-natural affection towards merging their curiosity towards solving real life-based problems.
We wanted to complete our work while having similar skeptical thoughts, if our knowledge in this module can help us to make proper contribution to our global or per say, the problematic consultations of humanity, it would be sentiment of pride and honor for we. What and why we proposed our case to do a job that comes with difficult handling and a larger parameter of understanding for computer vision. We hope to fulfill our perception of our ability to apply our knowledge in the betterment of our society.
What and why understanding these methods is crucial to our work on this research. Since the images vary, a large amount of contrast in the case of detection and segmentation accuracy panel.
Introduction
We also need to determine about the images we are willing to perform for this type of research.
Background Studies
Research Methodology
Experimental Results and Discussion
Implementation
Conclusion and Future Scope
Background Studies
- Related Work
- Research Summary
- Scope of the problem
- Challenges
And in recent works to be mentioned when it comes to our model, there was a zero-shot detection process if they were able to use an approach to train a joint embedding to acquire the query image and textual elaboration, while they skip visual elaboration in the novel categorization. In the past decade, computer vision has made its way to a lot of prosperity in the past decade, there have been consequential improvements in computer science. Computer vision mainly consists of the work of locating and individualizing a certain marked object based on visualized data [25].
26] where they used FCNN (Fully Convolutional Neural Network), consistently modifying the segmentation order of the roads. In such cases they added a tolerance rule in case of dealing with significant errors and in case of increasing the quality of road extraction. In their work they used Encoder and Decoder to understand the model towards training and getting the segmentation output stage.
By combining the close connection mode with the U-Net, they were able to overcome the problem of the occurrence of trees and shadows. And in the process to emphasize on the foreground pixels, they used a weighted loss function to work correctly. The authors improved FCNNs by adding a tolerance rule for spatially modest defects to extract roadways.
To address tree and shadow occlusion issues and highlight foreground pixels, the model mixes dense connectivity with U-Net. In order to use deep features for shallow feature maps with high-resolution photographs, shallow feature channels and deep feature channels in this study were combined. According to the authors, difficult classified samples with fewer pixels foreground should be taken into account by calculating the mask loss in the DL model using the focus loss function.
Compared to state-of-the-art approaches, the tests showed that the proposed method improved road marking segmentation. When we work in the field of computer vision, we find that when the data set is completely conclusive, the other parts are quite short-sighted to overcome. There are many consequences when the model appears to be confused with the classification process; more consciously, the model computes enough annotated parts and is unable to maintain phase while segmenting the role of the structural details of the road.
Research Methodology
- U-NET
- DeepLabV3+
- Research Subject and Instrumentation
- Road Dataset
- Dataset Distribution
- Preprocessing
- Statistical Analysis
- Implementation Requirements
- Model Performance
- Summary
Each road shape must fit into one of the design classifications, which is a polygon. 11 decoder levels in layer 4 of 21 there are cross-layer connections, which can help the up-sampling layer recover the image details. The U-net is proportional in design. By combining both the corresponding feature patterns of the encoder and decoder, segmentation is enabled by the detailed location information that can be stored more efficiently with shallow networks.
Many machine vision models use the neural network Res-Net (Residual Network), which serves as their foundation. Deeper neural network training is relatively difficult under normal circumstances.) It turns out that the vanishing gradient issue causes higher error rates while training deep neural networks. To obtain the relevant data, the knowledge area and global average aggregation of the feature map are used. The proposed model, DeepLabv3+, recovers the exact bounding boxes of the encoder modules, which also contributes rich semantic information to the model.
The collection includes flora, houses, roads, rivers and vehicles over a total area of more than 2,600 km2. The labels are binary images in which roads are displayed as constant 7-pixel thick lines generated by rasterizing and plotting Open Street Map's vector centerlines. Since computer vision is known to be the largest research sector, the annotation process is one of the most important works in the case of the work zone. After all the scrubbing and removing all the untamed data, the dataset was finally ready to be divided into classes for training and testing.
The entire amount of data was distributed in the most aformat way so that the models could perform as well as possible. Since computer vision is known to be the most extensive sector of the research field, the tagging process is one of the most important parts in the case of a work zone. Since our base model was adjusted, we also created a comparison table of measurements in case of obscuring the comparison of models.
The training loss in the figure is considered as the capacity of the model that consisted of the training data, and the validation loss indicated the ability of the model to understand the model's ability to handle the new data. It is proven that the final segmentation of the U-Net ResNet50 model is exactly as we expected. The segmentation via U-Net ResNet50 was more frequent and time-consuming than the rest of the models.
Indeed, one of the most fundamental techniques for comparing data samples in machine learning is the intersection over union (IOU) method. The models were able to perform according to our estimate of the performance result.
Impact on Society, Environment and Sustainability 5.1 Impact on Society
Impact on Environment
Ethical Aspects
Sustainability
Recommendations
Implication for Further Study
A dense pyramid network-based deep learning model for road marking instance segmentation using MLS point clouds. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention, Munich, Germany, 5-9. October 2015; Springer: Berlin, Germany, 2015; pp. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, Salt Lake City, UT, USA, 18-23. June 2018.