Wei Yuan

Prof. Wei Yuan

Chengdu University, Chengdu, Sichuan, China
orcid

Research Focus & Metrics

Academic Career Length

8Years (2019-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
30 September 2022 in Remote Sensing
Verified
Journal Article
Open access
PDF file
WYWei Yuan
Wei Yuan
JWJin Wang
Jin Wang
WXWenbo Xu
Wenbo Xu
Building extraction by deep learning from remote sensing images is currently a research hotspot. PSPNet is one of the classic semantic segmentation models and is currently adopted by many applications. Moreover, PSPNet can use not only CNN-based networks but also transformer-based networks as backbones; therefore, PSPNet also has high value in the transformer era. The core of PSPNet is the pyramid pooling module, which gives PSPNet the ability to capture the local features of different scales. However, the pyramid pooling module also has obvious shortcomings. The grid is fixed, and the pixels close to the edge of the grid cannot obtain the entire local features. To address this issue, an improved PSPNet network architecture named shift pooling PSPNet is proposed, which uses a module called shift pyramid pooling to replace the original pyramid pooling module, so that the pixels at the edge of the grid can also obtain the entire local features. Shift pooling is not only useful for PSPNet but also in any network that uses a fixed grid for downsampling to increase the receptive field and save computing, such as ResNet. A dense connection was adopted in decoding, and upsampling was gradually carried out. With two open datasets, the improved PSPNet, PSPNet, and some classic image segmentation models were used for comparative experiments. The results show that our method is the best according to the evaluation metrics, and the predicted image is closer to the label.

23 November 2021 in Remote Sensing
Verified
Journal Article
Open access
PDF file
WYWei Yuan
Wei Yuan
WXWenbo Xu
Wenbo Xu
The segmentation of remote sensing images by deep learning technology is the main method for remote sensing image interpretation. However, the segmentation model based on a convolutional neural network cannot capture the global features very well. A transformer, whose self-attention mechanism can supply each pixel with a global feature, makes up for the deficiency of the convolutional neural network. Therefore, a multi-scale adaptive segmentation network model (MSST-Net) based on a Swin Transformer is proposed in this paper. Firstly, a Swin Transformer is used as the backbone to encode the input image. Then, the feature maps of different levels are decoded separately. Thirdly, the convolution is used for fusion, so that the network can automatically learn the weight of the decoding results of each level. Finally, we adjust the channels to obtain the final prediction map by using the convolution with a kernel of 1 × 1. By comparing this with other segmentation network models on a WHU building data set, the evaluation metrics, mIoU, F1-score and accuracy are all improved. The network model proposed in this paper is a multi-scale adaptive network model that pays more attention to the global features for remote sensing segmentation.

20 May 2021 in IEEE Access
Verified
Journal Article
Open access
WYWei Yuan
Wei Yuan
WXWenbo Xu
Wenbo Xu
House segmentation of remote sensing image based on deep learning has become the main segmentation method because it can automatically extract features. However, the accuracy of image segmentation is affected not only by the network model, but also by the loss function, but the existing loss functions, except Binary Cross Entropy, are designed to deal with imbalanced dataset, no new research on improving Binary Cross Entropy for balanced dataset, and all loss function treat each pixel in isolation, without considering the spatial correlation between pixel and its neighbor pixels. To solve this problem, a new loss function, named NeighborLoss function, is proposed. Firstly, the deep learning network is used to get the prediction results of each pixel. According to whether the prediction results of the eight neighboring pixels of each pixel are consistent with the each pixel prediction, different weights are given to each pixel. Finally, the weighted average value of cross entropy of all pixels in the batch is taken as the final loss function value. We use the main deep learning semantic segmentation networks SegNet, PSPNet, UNET ++, MUNet with both NeighborLoss and cross entropy loss respectively to extract houses on the open data set named WHU dataset for remote sensing. The results show that compared with cross entropy loss functions, the MIoU, Precision, Recall, and Accuracy of NeighborLoss function are improved. From the predicted graph, the NeighborLoss function is more accurate to extract the edge of the house, especially in the corner of the house. NeighborLoss function is a more effective loss function for remote sensing image segmentation.

10 April 2023 in Remote Sensing
Verified
Journal Article
Open access
PDF file
WYWei Yuan
Wei Yuan
XZXiaobo Zhang
Xiaobo Zhang
JSJibao Shi
Jibao Shi
JWJin Wang
Jin Wang
Extracting building data from remote sensing images is an efficient way to obtain geographic information data, especially following the emergence of deep learning technology, which results in the automatic extraction of building data from remote sensing images becoming increasingly accurate. A CNN (convolution neural network) is a successful structure after a fully connected network. It has the characteristics of saving computation and translation invariance with improved local features, but it has difficulty obtaining global features. Transformers can compensate for the shortcomings of CNNs and more effectively obtain global features. However, the calculation number of transformers is excessive. To solve this problem, a Lite Swin transformer is proposed. The three matrices Q, K, and V of the transformer are simplified to only a V matrix, and the v of the pixel is then replaced by the v with the largest projection value on the pixel feature vector. In order to better integrate global features and local features, we propose the LiteST-Net model, in which the features extracted by the Lite Swin transformer and the CNN are added together and then sampled up step by step to fully utilize the global feature acquisition ability of the transformer and the local feature acquisition ability of the CNN. The comparison experiments on two open datasets are carried out using our proposed LiteST-Net and some classical image segmentation models. The results show that compared with other networks, all metrics of LiteST-Net are the best, and the predicted image is closer to the label.

18 May 2022 in Remote Sensing
Verified
Journal Article
Open access
PDF file
WYWei Yuan
Wei Yuan
WXWenbo Xu
Wenbo Xu
At present, road continuity is a major challenge, and it is difficult to extract the centerline vector of roads, especially when the road view is obstructed by trees or other structures. Most of the existing research has focused on optimizing the available deep-learning networks. However, the segmentation accuracy is also affected by the loss function. Currently, little research has been published on road segmentation loss functions. To resolve this problem, an attention loss function named GapLoss that can be combined with any segmentation network was proposed. Firstly, a deep-learning network was used to obtain a binary prediction mask. Secondly, a vector skeleton was extracted from the prediction mask. Thirdly, for each pixel, eight neighboring pixels with the same value of the pixel were calculated. If the value was 1, then the pixel was identified as the endpoint. Fourth, according to the number of endpoints within a buffered range, each pixel in the prediction image was given a corresponding weight. Finally, the weighted average value of the cross-entropy of all the pixels in the batch was used as the final loss function value. We employed four well-known semantic segmentation networks to conduct comparative experiments on three large datasets. The results showed that, compared to other loss functions, the evaluation metrics after using GapLoss were nearly all improved. From the predicted image, the road prediction by GapLoss was more continuous, especially at intersections and when the road was obscured from view, and the road segmentation accuracy was improved.

1 August 2024 in Forests
Verified
Journal Article
Open access
PDF file
WYWei Yuan
Wei Yuan
磊乔磊 乔
磊 乔
LTLiu Tang
Liu Tang
Cameras, especially those carried by drones, are the main tools used to detect wildfires in forests because cameras have much longer detection ranges than smoke sensors. Currently, deep learning is main method used for fire detection in images, and Transformer is the best algorithm. Swin Transformer restricts the computation to a fixed-size window, which reduces the amount of computation to a certain extent, but to allow pixel communication between windows, it adopts a shift window approach. Therefore, Swin Transformer requires multiple shifts to extend the receptive field to the entire image. This somewhat limits the network’s ability to capture global features at different scales. To solve this problem, instead of using the shift window method to allow pixel communication between windows, we downsample the feature map to the window size after capturing global features through a single Transformer, and we upsample the feature map to the original size and add it to the previous feature map. This way, there is no need for multiple layers of stacked window Transformers; global features are captured after each window Transformer operation. We conducted experiments on the Corsican fire dataset captured by ground cameras and on the Flame dataset captured by drone cameras. The results show that our algorithm performs the best. On the Corsican fire dataset, the mIoU, F1 score, and OA reached 79.4%, 76.6%, and 96.9%, respectively. On the Flame dataset, the mIoU, F1 score, and OA reached 84.4%, 81.6%, and 99.9%, respectively.