Spatial Pyramid Attention for Deep Convolutional Neural Networks

Ma, Xu; Guo, Jingda; Sansom, Andrew; McGuire, Mara; Kalaani, Andrew; Chen, Qi; Tang, Sihai; Yang, Qing; Fu, Song

doi:10.1109/tmm.2021.3068576

Cited by 29 publications

(9 citation statements)

References 26 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The SPA is originally proposed in the previous work (Ma et al 2021). It adopts a spatial pyramid structure to characterize three-scale attention weights.…”

Section: Cross-level Feature Fusion Modulementioning

confidence: 99%

See 1 more Smart Citation

Multi-scale feature aggregation and fusion network with self-supervised multi-level perceptual loss for textures preserving low-dose CT denoising

Zhang,

Wan,

Wang

et al. 2024

Phys. Med. Biol.

View full text Add to dashboard Cite

Objective. The textures and detailed structures in computed tomography (CT) images are highly desirable for clinical diagnosis. This study aims to expand the current body of work on textures and details preserving convolutional neural networks for low-dose CT (LDCT) image denoising task. Approach. This study proposed a novel Multi-scale Feature Aggregation and Fusion network (MFAF-net) for LDCT image denoising. Specifically, we proposed a Multi-scale Residual Feature Aggregation Module (MRFAM) to characterize multi-scale structural information in CT images, which captures regional-specific inter-scale variations using learned weights. We further proposed a Cross-level Feature Fusion Module (CFFM) to integrate cross-level features, which adaptively weights the contributions of features from encoder to decoder by using a Spatial Pyramid Attention (SPA) mechanism. Moreover, we proposed a Self-supervised Multi-level Perceptual Loss Module (SMPLM) to generate multi-level auxiliary perceptual supervision for recovery of salient textures and structures of tissues and lesions in CT images, which takes advantage of abundant semantic information at various levels. We introduced parameters for the perceptual loss to adaptively weight the contributions of auxiliary features of different levels and we also introduced an automatic parameter tuning strategy for these parameters. Main results. Extensive experimental studies were performed to validate the effectiveness of the proposed method. Experimental results demonstrate that the proposed method can achieve better performance on both fine textures preservation and noise suppression for CT image denoising task compared with other competitive CNN based methods. Significance. The proposed MFAF-net takes advantage of multi-scale receptive fields, cross-level features integration and self-supervised multi-level perceptual loss, enabling more effective recovering of fine textures and detailed structures of tissues and lesions in CT images.

show abstract

“…The SPA is originally proposed in the previous work (Ma et al 2021). It adopts a spatial pyramid structure to characterize three-scale attention weights.…”

Section: Cross-level Feature Fusion Modulementioning

confidence: 99%

“…In this study, we extend the SPA structure by aggregate the feature context at four different scales for better capturing fine, local, coarse, and global contextual information in CT images, as illustrated in figure 3(b). For conserving space, please refer to Ma et al (2021) for detail information of the SPA structure.…”

Section: Cross-level Feature Fusion Modulementioning

confidence: 99%

Multi-scale feature aggregation and fusion network with self-supervised multi-level perceptual loss for textures preserving low-dose CT denoising

Zhang,

Wan,

Wang

et al. 2024

Phys. Med. Biol.

View full text Add to dashboard Cite

show abstract

“…The Thirty-Eighth AAAI Conference on Artificial Intelligence (AAAI-24) information and are capable of retaining the spatial characteristics within each channel (Ma et al 2021). As shown in Figure 4, we use multiple max-pooling layers to eliminate redundant spatial information in the feature maps.…”

Section: Cyclic Disentanglement Modulementioning

confidence: 99%

Triple Feature Disentanglement for One-Stage Adaptive Object Detection

Wang,

Jia,

Zeng

et al. 2024

AAAI

View full text Add to dashboard Cite

In recent advancements concerning Domain Adaptive Object Detection (DAOD), unsupervised domain adaptation techniques have proven instrumental. These methods enable enhanced detection capabilities within unlabeled target domains by mitigating distribution differences between source and target domains. A subset of DAOD methods employs disentangled learning to segregate Domain-Specific Representations (DSR) and Domain-Invariant Representations (DIR), with ultimate predictions relying on the latter. Current practices in disentanglement, however, often lead to DIR containing residual domain-specific information. To address this, we introduce the Multi-level Disentanglement Module (MDM) that progressively disentangles DIR, enhancing comprehensive disentanglement. Additionally, our proposed Cyclic Disentanglement Module (CDM) facilitates DSR separation. To refine the process further, we employ the Categorical Features Disentanglement Module (CFDM) to isolate DIR and DSR, coupled with category alignment across scales for improved source-target domain alignment. Given its practical suitability, our model is constructed upon the foundational framework of the Single Shot MultiBox Detector (SSD), which is a one-stage object detection approach. Experimental validation highlights the effectiveness of our method, demonstrating its state-of-the-art performance across three benchmark datasets.

show abstract

“…Fully convolutional network (FCN) was effectively used for semantic segmentation [1] and, consequently, many deep learning-based methods were proposed. They can be mainly divided into three categories: gradually recover resolution-based methods [4], multi-scale pyramid-based methods [19][20][21] and attention module-based methods [22][23][24]. The U-shaped network was proposed in Ref.…”

Section: Introductionmentioning

confidence: 99%

Refinement Co‐supervision network for real‐time semantic segmentation

Dong

Zhao

Zheng

et al. 2023

IET Computer Vision

View full text Add to dashboard Cite

Semantic segmentation is a fundamental technology for autonomous driving. It has a high demand for inference speed and accuracy. However, a good trade‐off between accuracy and latency is yet not present in existing semantic segmentation approaches. Due to the limitation of speed, the authors cannot increase the number of network layers without limit and cannot design modules like in the networks without real‐time. It is a challenging problem how to design a model with good performance under limited resources. To alleviate these issues, in this study, the authors propose a refinement co‐supervision network (RCNet), which is real‐time on a high‐resolution image (1024×2048). The authors first construct the context refinement module, which can provide low computation cost way for obtaining the large receptive field and context information. Furthermore, a boundary co‐supervision mechanism is proposed. It strengthens the optimisation of easily neglected boundaries and small targets. Experimental results reveal that the proposed RCNet outperforms seven representative semantic segmentation methods.

show abstract

Spatial Pyramid Attention for Deep Convolutional Neural Networks

Cited by 29 publications

References 26 publications

Multi-scale feature aggregation and fusion network with self-supervised multi-level perceptual loss for textures preserving low-dose CT denoising

Multi-scale feature aggregation and fusion network with self-supervised multi-level perceptual loss for textures preserving low-dose CT denoising

Triple Feature Disentanglement for One-Stage Adaptive Object Detection

Refinement Co‐supervision network for real‐time semantic segmentation

Contact Info

Product

Resources

About