Zeji Wang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Zeji Wang has a total of 6 co-authors. Here are the top co-authors by number of shared publications:
- Xiaowei He
School of Computer Science and Technology, Zhejiang Normal University, Jinhua, China
2 shared publications - Yi Li
College of Mathematics and Computer Science, Zhejiang Normal University, Jinhua 321004, China
2 shared publications - Lingna Wang
College of Mathematics and Computer Science, Zhejiang Normal University, Jinhua 321004, China
1 shared publication - Qinliang Chuai
School of Computer Science and Technology, Zhejiang Normal University, Jinhua 321004, China
1 shared publication - Rao Cheng
College of Mathematics and Computer Science, Zhejiang Normal University, Jinhua 321004, China
1 shared publication - Zhonglong Zheng
Zhejiang Normal University
1 shared publication
Latest Publications
- Open Access
EmbedFormer: Embedded Depth-Wise Convolution Layer for Token Mixing
Visual Transformers (ViTs) have shown impressive performance due to their powerful coding ability to catch spatial and channel information. MetaFormer gives us a general architecture of transformers consisting of a token mixer and a channel mixer through which we can generally understand how transformers work. It is proved that the general architecture of the ViTs is more essential to the models’ performance than self-attention mechanism. Then, Depth-wise Convolution layer (DwConv) is widely accepted to replace local self-attention in transformers. In this work, a pure convolutional "transformer" is designed. We rethink the difference between the operation of self-attention and DwConv. It is found that the self-attention layer, with an embedding layer, unavoidably affects channel information, while DwConv only mixes the token information per channel. To address the differences between DwConv and self-attention, we implement DwConv with an embedding layer before as the token mixer to instantiate a MetaFormer block and a model named EmbedFormer is introduced. Meanwhile, SEBlock is applied in the channel mixer part to improve performance. On the ImageNet-1K classification task, EmbedFormer achieves top-1 accuracy of 81.7% without additional training images, surpassing the Swin transformer by +0.4% in similar complexity. In addition, EmbedFormer is evaluated in downstream tasks and the results are entirely above those of PoolFormer, ResNet and DeiT. Compared with PoolFormer-S24, another instance of MetaFormer, our EmbedFormer improves the score by +3.0% box AP/+2.3% mask AP on the COCO dataset and +1.3% mIoU on the ADE20K.
Open AccessEmbedFormer: Embedded Depth-Wise Convolution Layer for Token Mixing
Visual Transformers (ViTs) have shown impressive performance due to their powerful coding ability to catch spatial and channel information. MetaFormer gives us a general architecture of transformers consisting of a token mixer and a channel mixer through which we can generally understand how transformers work. It is proved that the general architecture of the ViTs is more essential to the models’ performance than self-attention mechanism. Then, Depth-wise Convolution layer (DwConv) is widely accepted to replace local self-attention in transformers. In this work, a pure convolutional "transformer" is designed. We rethink the difference between the operation of self-attention and DwConv. It is found that the self-attention layer, with an embedding layer, unavoidably affects channel information, while DwConv only mixes the token information per channel. To address the differences between DwConv and self-attention, we implement DwConv with an embedding layer before as the token mixer to instantiate a MetaFormer block and a model named EmbedFormer is introduced. Meanwhile, SEBlock is applied in the channel mixer part to improve performance. On the ImageNet-1K classification task, EmbedFormer achieves top-1 accuracy of 81.7% without additional training images, surpassing the Swin transformer by +0.4% in similar complexity. In addition, EmbedFormer is evaluated in downstream tasks and the results are entirely above those of PoolFormer, ResNet and DeiT. Compared with PoolFormer-S24, another instance of MetaFormer, our EmbedFormer improves the score by +3.0% box AP/+2.3% mask AP on the COCO dataset and +1.3% mIoU on the ADE20K.
- Open Access
Single-Shot Object Detection via Feature Enhancement and Channel Attention
Features play a critical role in computer vision tasks. Deep learning methods have resulted in significant breakthroughs in the field of object detection, but it is still an extremely challenging obstacle when an object is very small. In this work, we propose a feature-enhancement- and channel-attention-guided single-shot detector called the FCSSD with four modules to improve object detection performance. Specifically, inspired by the structure of atrous convolution, we built an efficient feature-extraction module (EFM) in order to explore contextual information along the spatial dimension, and then pyramidal aggregation module (PAM) is presented to explore the semantic features of deep layers, thus reducing the semantic gap between multi-scale features. Furthermore, we construct an effective feature pyramid refinement fusion (FPRF) to refine the multi-scale features and create benefits for richer object knowledge. Finally, an attention-guided module (AGM) is developed to balance the channel weights and optimize the final integrated features on each level; this alleviates the aliasing effects of the FPN with negligible computational costs. The FCSSD exploits richer information of shallow layers and higher layers by using our designed modules, thus accomplishing excellent detection performance for multi-scale object detection and reaching a better tradeoff between accuracy and inference time. Experiments on PASCAL VOC and MS COCO datasets were conducted to evaluate the performance, showing that our FCSSD achieves competitive detection performance compared with existing mainstream object detection methods.
Open AccessSingle-Shot Object Detection via Feature Enhancement and Channel Attention
Features play a critical role in computer vision tasks. Deep learning methods have resulted in significant breakthroughs in the field of object detection, but it is still an extremely challenging obstacle when an object is very small. In this work, we propose a feature-enhancement- and channel-attention-guided single-shot detector called the FCSSD with four modules to improve object detection performance. Specifically, inspired by the structure of atrous convolution, we built an efficient feature-extraction module (EFM) in order to explore contextual information along the spatial dimension, and then pyramidal aggregation module (PAM) is presented to explore the semantic features of deep layers, thus reducing the semantic gap between multi-scale features. Furthermore, we construct an effective feature pyramid refinement fusion (FPRF) to refine the multi-scale features and create benefits for richer object knowledge. Finally, an attention-guided module (AGM) is developed to balance the channel weights and optimize the final integrated features on each level; this alleviates the aliasing effects of the FPN with negligible computational costs. The FCSSD exploits richer information of shallow layers and higher layers by using our designed modules, thus accomplishing excellent detection performance for multi-scale object detection and reaching a better tradeoff between accuracy and inference time. Experiments on PASCAL VOC and MS COCO datasets were conducted to evaluate the performance, showing that our FCSSD achieves competitive detection performance compared with existing mainstream object detection methods.
- Open Access
Small Object Detection in Traffic Scenes Based on YOLO-MXANet
In terms of small objects in traffic scenes, general object detection algorithms have low detection accuracy, high model complexity, and slow detection speed. To solve the above problems, an improved algorithm (named YOLO-MXANet) is proposed in this paper. Complete-Intersection over Union (CIoU) is utilized to improve loss function for promoting the positioning accuracy of the small object. In order to reduce the complexity of the model, we present a lightweight yet powerful backbone network (named SA-MobileNeXt) that incorporates channel and spatial attention. Our approach can extract expressive features more effectively by applying the Shuffle Channel and Spatial Attention (SCSA) module into the SandGlass Block (SGBlock) module while increasing the parameters by a small number. In addition, the data enhancement method combining Mosaic and Mixup is employed to improve the robustness of the training model. The Multi-scale Feature Enhancement Fusion (MFEF) network is proposed to fuse the extracted features better. In addition, the SiLU activation function is utilized to optimize the Convolution-Batchnorm-Leaky ReLU (CBL) module and the SGBlock module to accelerate the convergence of the model. The ablation experiments on the KITTI dataset show that each improved method is effective. The improved algorithm reduces the complexity and detection speed of the model while improving the object detection accuracy. The comparative experiments on the KITTY dataset and CCTSDB dataset with other algorithms show that our algorithm also has certain advantages.
Open AccessSmall Object Detection in Traffic Scenes Based on YOLO-MXANet
In terms of small objects in traffic scenes, general object detection algorithms have low detection accuracy, high model complexity, and slow detection speed. To solve the above problems, an improved algorithm (named YOLO-MXANet) is proposed in this paper. Complete-Intersection over Union (CIoU) is utilized to improve loss function for promoting the positioning accuracy of the small object. In order to reduce the complexity of the model, we present a lightweight yet powerful backbone network (named SA-MobileNeXt) that incorporates channel and spatial attention. Our approach can extract expressive features more effectively by applying the Shuffle Channel and Spatial Attention (SCSA) module into the SandGlass Block (SGBlock) module while increasing the parameters by a small number. In addition, the data enhancement method combining Mosaic and Mixup is employed to improve the robustness of the training model. The Multi-scale Feature Enhancement Fusion (MFEF) network is proposed to fuse the extracted features better. In addition, the SiLU activation function is utilized to optimize the Convolution-Batchnorm-Leaky ReLU (CBL) module and the SGBlock module to accelerate the convergence of the model. The ablation experiments on the KITTI dataset show that each improved method is effective. The improved algorithm reduces the complexity and detection speed of the model while improving the object detection accuracy. The comparative experiments on the KITTY dataset and CCTSDB dataset with other algorithms show that our algorithm also has certain advantages.