Efficient Depth Fusion Transformer for Aerial Image Semantic Segmentation

Li, Yan; Huang, Jianming; Xie, Hong; Wei, Pengcheng; Zhao, Guoqing

doi:10.3390/rs14051294

Cited by 15 publications

(3 citation statements)

References 40 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…This basic example shows that after going through the patch merging layer, the height and width of the feature map are halved, but the depth is doubled. To reduce the resolution, adjust the number of channels used, and complete the hierarchical design of the ST, merge correction was used [14]. In this case, each downsample is two, and elements are chosen at every other point in the row and column directions before being spliced together to expand, as seen in Figure 8.…”

Section: Figure 3: the Architecture Of The Swin Transformermentioning

confidence: 99%

Implementation Of The Swin Transformer and Its Application In Image Classification

A. Dihin,

Al Shemmary,

Al-Jawher

2023

J.port.sci.res.

View full text Add to dashboard Cite

There are big differences between the field of view of the calculator and the field of natural languages, for example, in the field of vision, the difference is in the size of the object as well as in the accuracy of the pixels in the image, and this contradicts the words in the text, and this makes the adaptation of the transformers to see somewhat difficult.Very recently a vision transformer named Swin Transformer was introduced by the Microsoft research team in Asia to achieve state-of-the-art results for machine translation. The computational complexity is linear and proportional to the size of the input image, because the processing of subjective attention is within each local window separately, and thus results in processor maps that are hierarchical and in deeper layers, and thus serve as the backbone of the calculator's vision in image classification and dense recognition applications. This work focuses on applying the Swin transformer to a demonstrated mathematical example with step-by-step analysis. Additionally, extensive experimental results were carried out on several standardized databases from CIFAR-10, CIFAR-100, and MNIST. Their results showed that the Swin Transformer can achieve flexible memory savings. Test accuracy for CIFAR-10 gave a 71.54% score, while for the CIFAR-100 dataset the accuracy was 46.1%. Similarly, when the Swin transformer was applied to the MNIST dataset, the accuracy increased in comparison with other vision transformer results.

show abstract

Section: Figure 3: the Architecture Of The Swin Transformermentioning

confidence: 99%

Implementation Of The Swin Transformer and Its Application In Image Classification

A. Dihin,

Al Shemmary,

Al-Jawher

2023

J.port.sci.res.

View full text Add to dashboard Cite

show abstract

“…Some al-ternative approaches to image segmentation employ graph neural networks (GNN) [52][53][54], which operate on graph nodes constructed from an image in a preprocessing step. Another recent line of research in remote sensing image segmentation uses attention-based transformers to replace or supplement the backbone CNN [55][56][57][58].…”

Section: Introductionmentioning

confidence: 99%

Detection and Monitoring of Woody Vegetation Landscape Features Using Periodic Aerial Photography

et al. 2023

View full text Add to dashboard Cite

Woody vegetation landscape features, such as hedges, tree patches, and riparian vegetation, are important elements of landscape and biotic diversity. For the reason that biodiversity loss is one of the major ecological problems in the EU, it is necessary to establish efficient workflows for the registration and monitoring of woody vegetation landscape features. In the paper, we propose and evaluate a methodology for automated detection of changes in woody vegetation landscape features from a digital orthophoto (DOP). We demonstrate its ability to capture most of the actual changes in the field and thereby provide valuable support for more efficient maintenance of landscape feature layers, which is important for the shaping of future environmental policies. While the most reliable source for vegetation cover mapping is a combination of LiDAR and high-resolution imagery, it can be prohibitively expensive for continuous updates. The DOP from cyclic aerial photography presents an alternative source of up-to-date information for tracking woody vegetation landscape features in-between LiDAR recordings. The proposed methodology uses a segmentation neural network, which is trained with the latest DOP against the last known ground truth as the target. The output is a layer of detected changes, which are validated by the user before being used to update the woody vegetation landscape feature layer. The methodology was tested using the data of a typical traditional Central European cultural landscape, Goričko, in north-eastern Slovenia. The achieved F1 of per-pixel segmentation was 83.5% and 77.1% for two- and five-year differences between the LiDAR-based reference and the DOP, respectively. The validation of the proposed changes at a minimum area threshold of 100 m2 and a minimum area percentage threshold of 20% showed that the model achieved recall close to 90%.

show abstract

“…Many scholars have attempted to improve segmentation accuracy by using deep convolutional networks and feature fusion operations. Yan et al [21] proposed the fusion of color and depth images at the front end of feature extraction to improve segmentation accuracy. Wang et al [22]proposed further improvement of segmentation accuracy by combining the high-level features of two different images obtained by parallel deep networks.…”

Section: Introductionmentioning

confidence: 99%

CM-Unet: A Novel Remote Sensing Image Segmentation Method Based on Improved U-Net

Cui,

Li,

Chen

et al. 2023

IEEE Access

View full text Add to dashboard Cite

Semantic segmentation is an active research area for high-resolution (HR) remote sensing image processing. Most existing algorithms are better at segmenting different features. However, for complex scenes, many algorithms have insufficient segmentation accuracy. In this study, we propose a new method CM-Unet based on the U-Net framework to address the problems of holes, omissions, and fuzzy edge segmentation. First, we add the channel attention mechanism in the encoding network and the residual module to transmit information. Second, a multi-feature fusion mechanism is proposed in the decoding network, and an improved sub-pixel convolution method replaces the traditional upsampling operation. We conducted simulation experiments on the Potsdam, Vaihingen and GID datasets. The experimental results show that the proposed CM-Unet required segmentation time is approximately 62 ms/piece, the MIoU is 90.4% and the floating point operations (FLOPs) is 20.95 MFLOPs. Compared with U-Net, CM-Unet only increased the total number of parameters and floating point operations slightly, but achieved the best segmentation effect compared with the other models. CM-Unet can segment remote sensing images efficiently and accurately owing to its lower time consumption and space requirements; the precision of the segmentation results is better than other methods.

show abstract

Efficient Depth Fusion Transformer for Aerial Image Semantic Segmentation

Cited by 15 publications

References 40 publications

Implementation Of The Swin Transformer and Its Application In Image Classification

Implementation Of The Swin Transformer and Its Application In Image Classification

Detection and Monitoring of Woody Vegetation Landscape Features Using Periodic Aerial Photography

CM-Unet: A Novel Remote Sensing Image Segmentation Method Based on Improved U-Net

Contact Info

Product

Resources

About