Transfer learning and U-Net for buildings segmentation

Adiba, Abdelilah; Hajji, Hicham; Maâtouk, Mustapha

doi:10.1145/3314074.3314088

Cited by 12 publications

(11 citation statements)

References 10 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We choose Dice Coefficient as evaluation metric, also known as IoU, which is the same as our baseline (U-Net ResNet-34) [1]. It is very similar to the jaccard coefficient, but more accurate, which can be defined as:…”

Section: Metricsmentioning

confidence: 99%

“…We compare CT-UNet with the latest encoder-decoder methods on the Inria Aerial Image Labeling Dataset. SegNet (Multi-Task Loss) [5], 2-levels U-Nets [13] and U-Net ResNet-34 [1] are used as reference. We calculate the overall accuracy and mean IoU in the validation set, shown the quantitative results in Table 3.…”

Section: Inria Aerial Image Labeling Datasetmentioning

confidence: 99%

“…Sebastian et al [26] propose a novel loss function combining both adversarial and cross-entropy losses. U-Net ResNet-34 [1] replaces the decoder of U-Net with ResNet to achieve better segmentation performance on buildings. This proves that U-Net has a good effect on building segmentation.…”

Section: Introductionmentioning

confidence: 99%

“…However, the current automatic segmentation tasks for buildings still have the following problems. (1) The high resolution and large size of remote sensing images lead to most networks exhibit poor recognition capabilities. This causes the outlines of the segmented building maps are blurred.…”

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

CT-UNet: Context-Transfer-UNet for Building Segmentation in Remote Sensing Images

Liu

Jin

et al. 2021

Neural Process Lett

View full text Add to dashboard Cite

With the proliferation of remote sensing images, how to segment buildings more accurately in remote sensing images is a critical challenge. First, most networks have poor recognition ability on high resolution images, resulting in blurred boundaries in the segmented building maps. Second, the similarity between buildings and background results in intraclass inconsistency. To address these two problems, we propose an UNet-based network named Context-Transfer-UNet (CT-UNet). Specifically, we design Dense Boundary Block. Dense Block utilizes reuse mechanism to refine features and increase recognition capabilities. Boundary Block introduces the low-level spatial information to solve the fuzzy boundary problem. Then, to handle intra-class inconsistency, we construct Spatial Channel Attention Block. It combines context space information and selects more distinguishable features from space and channel. Finally, we propose an improved loss function to enhance the purpose of loss by adding evaluation indicator. Based on our proposed CT-UNet, we achieve 85.33% mean IoU on the Inria dataset, 91.00% mean IoU on the WHU dataset and 83.92% F1-score on the Massachusetts dataset. The results outperform our baseline (U-Net ResNet-34) by 3.76%, exceed Web-Net by 2.24% and surpass HFSA-Unet by 2.17%.

show abstract

Section: Metricsmentioning

confidence: 99%

Section: Inria Aerial Image Labeling Datasetmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

CT-UNet: Context-Transfer-UNet for Building Segmentation in Remote Sensing Images

Liu

Jin

et al. 2021

Neural Process Lett

View full text Add to dashboard Cite

show abstract

“…While Cheetah's use of data augmentation reduces the amount of manually annotated images required for training, there is growing interest in the area of transfer learning where a fully trained artificial neural network can be quickly adapted for effective use in a new, but related task 40 . Given that other U-Net based segmentation systems have demonstrated the ability for transfer learning in other fields 41 , this would be an interesting feature for future implementation in Cheetah.…”

Section: Discussionmentioning

confidence: 99%

Cheetah: a computational toolkit for cybergenetic control

Pedone

Cesare

Zamora

et al. 2020

Preprint

View full text Add to dashboard Cite

1Advances in microscopy, microfluidics and optogenetics enable single cell monitoring and 2 environmental regulation and offer the means to control cellular phenotypes. The development 3 of such systems is challenging and often results in bespoke setups that hinder reproducibility. To 4 address this, we introduce Cheetah -a flexible computational toolkit that simplifies the integration 5 of real-time microscopy analysis with algorithms for cellular control. Central to the platform is an 6 image segmentation system based on the versatile U-Net convolutional neural network. This is 7 supplemented with functionality to robustly count, characterise and control cells over time. We 8 demonstrate Cheetah's core capabilities by analysing long-term bacterial and mammalian cell 9 growth and by dynamically controlling protein expression in mammalian cells. In all cases, 10 Cheetah's segmentation accuracy exceeds that of a commonly used thresholding-based method, 11 allowing for more accurate control signals to be generated. Availability of this easy-to-use 12 platform will make control engineering techniques more accessible and offer new ways to probe 13 and manipulate living cells. 14 Introduction 15Modern automated microscopy techniques enable researchers to collect vast amounts of single-16 cell imaging data at high temporal resolutions. This has resulted in time-lapse microscopy 17 becoming the go to method for studying cellular dynamics, enabling the quantification of 18 processes such as stochastic fluctuations during gene expression 1-3 , emerging oscillatory 19 patterns in protein concentrations 4 , lineage selection 5,6 , and many more 7 . 20To make sense of microscopy images, segmentation is performed whereby an image is 21 broken up into regions corresponding to specific features of interest (e.g. cells and the 22 background). Image segmentation allows for the accurate quantification of cellular phenotypes 23 encoded by visual cues (e.g. fluorescence) by ensuring only those pixels corresponding to a cell 24 are considered. A range of segmentation algorithms have been proposed to automatically 25 analyse images of various organisms and tissues 3,8-11 . The most common of these are 26 thresholding 12 and seeded watershed 13 methods, which are available in many scientific image 27 processing toolkits. Commercial software packages also implement this type of functionality, 28 enabling both automated image acquisition and analysis (e.g. NIS-Elements, Nikon). While these 29 proprietary systems are user-friendly requiring no programming skills to be used, they are often 30 difficult to tailor for specific needs and cannot be easily extended to new forms of analysis. 31More recently, deep learning-based approaches to image segmentation have emerged 32 7,14-17 . Compared to the more common thresholding-based approaches 12 , deep learning methods 33 tend to require more significant computational resources when running on traditional computer 34 architectures and often require the time-consuming manual step of generating...

show abstract