Squeeze-and-Attention Networks for Semantic Segmentation

Zhong, Zilong; Lin, Zhong Qiu; Bidart, Rene; Hu, Xiaodan; Daya, Ibrahim Ben; Li, Zhifeng; Zheng, Wei‐Shi; Li, Jonathan; Wong, Alexander

doi:10.48550/arxiv.1909.03402

Cited by 4 publications

(2 citation statements)

References 49 publications

(76 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…It dynamically restructures the networks for the sake of different receptive fields in dilated convolutions [34,35]. In semantic segmentation, [36] imposes a pixel-group attention to remedy the deficiency of spatial information in SENet and [15] builds a link between each pixel and its surrounding pixels to capture important information. Attention mechanism is designed to dynamically calibrate the information flow in the forward propagation by learnable method.…”

Section: Related Workmentioning

confidence: 99%

Dynamic Region-Aware Convolution

Chen¹,

Wang²,

Guo³

et al. 2020

Preprint

View full text Add to dashboard Cite

We propose a new convolution called Dynamic Region-Aware Convolution (DRConv), which can automatically assign multiple filters to corresponding spatial regions where features have similar representation. In this way, DRConv outperforms standard convolution in modeling semantic variations. Standard convolution can increase the number of channels to extract more visual elements but results in high computational cost. More gracefully, our DRConv transfers the increasing channel-wise filters to spatial dimension with learnable instructor, which significantly improves representation ability of convolution and maintains translation-invariance like standard convolution. DRConv is an effective and elegant method for handling complex and variable spatial information distribution. It can substitute standard convolution in any existing networks for its plug-and-play property. We evaluate DRConv on a wide range of models (MobileNet series, ShuffleNetV2, etc.) and tasks (Classification, Face Recognition, Detection and Segmentation.). On ImageNet classification, DRConv-based ShuffleNetV2-0.5× achieves state-of-the-art performance of 67.1% at 46M multiply-adds level with 6.3% relative improvement.

show abstract

Section: Related Workmentioning

confidence: 99%

Dynamic Region-Aware Convolution

Chen¹,

Wang²,

Guo³

et al. 2020

Preprint

View full text Add to dashboard Cite

show abstract

“…Deep Learning based neural networks have provided tremendous improvements over the past decade in various domains of Computer Vision. These include Image Classification [24,17,20,40], Object Detection [31,3,26,30,12], Semantic Segmentation [16,4,44,5,45] among others. The drawbacks of these methods however include the fact that a large amount of computational resources are required to achieve state-of-the-art accuracy.…”

Section: Introductionmentioning

confidence: 99%

Online Ensemble Model Compression Using Knowledge Distillation

Walawalkar

Shen

Savvides

2020

Lecture Notes in Computer Science

View full text Add to dashboard Cite

This paper presents a novel knowledge distillation based model compression framework consisting of a student ensemble. It enables distillation of simultaneously learnt ensemble knowledge onto each of the compressed student models. Each model learns unique representations from the data distribution due to its distinct architecture. This helps the ensemble generalize better by combining every model's knowledge. The distilled students and ensemble teacher are trained simultaneously without requiring any pretrained weights. Moreover, our proposed method can deliver multi-compressed students with single training, which is efficient and flexible for different scenarios. We provide comprehensive experiments using state-of-the-art classification models to validate our framework's effectiveness. Notably, using our framework a 97% compressed ResNet110 student model managed to produce a 10.64% relative accuracy gain over its individual baseline training on CIFAR100 dataset. Similarly a 95% compressed DenseNet-BC (k=12) model managed a 8.17% relative accuracy gain.

show abstract

Automatic segmentation of fat metaplasia on sacroiliac joint MRI using deep learning

Li,

Lin,

Xie

et al. 2024

Insights Imaging

View full text Add to dashboard Cite

Objective To develop a deep learning (DL) model for segmenting fat metaplasia (FM) on sacroiliac joint (SIJ) MRI and further develop a DL model for classifying axial spondyloarthritis (axSpA) and non-axSpA. Materials and methods This study retrospectively collected 706 patients with FM who underwent SIJ MRI from center 1 (462 axSpA and 186 non-axSpA) and center 2 (37 axSpA and 21 non-axSpA). Patients from center 1 were divided into the training, validation, and internal test sets (n = 455, 64, and 129). Patients from center 2 were used as the external test set. We developed a UNet-based model to segment FM. Based on segmentation results, a classification model was built to distinguish axSpA and non-axSpA. Dice Similarity Coefficients (DSC) and area under the curve (AUC) were used for model evaluation. Radiologists’ performance without and with model assistance was compared to assess the clinical utility of the models. Results Our segmentation model achieved satisfactory DSC of 81.86% ± 1.55% and 85.44% ± 6.09% on the internal cross-validation and external test sets. The classification model yielded AUCs of 0.876 (95% CI: 0.811–0.942) and 0.799 (95% CI: 0.696–0.902) on the internal and external test sets, respectively. With model assistance, segmentation performance was improved for the radiological resident (DSC, 75.70% vs. 82.87%, p < 0.05) and expert radiologist (DSC, 85.03% vs. 85.74%, p > 0.05). Conclusions DL is a novel method for automatic and accurate segmentation of FM on SIJ MRI and can effectively increase radiologist’s performance, which might assist in improving diagnosis and progression of axSpA. Critical relevance statement DL models allowed automatic and accurate segmentation of FM on sacroiliac joint MRI, which might facilitate quantitative analysis of FM and have the potential to improve diagnosis and prognosis of axSpA. Key points • Deep learning was used for automatic segmentation of fat metaplasia on MRI. • UNet-based models achieved automatic and accurate segmentation of fat metaplasia. • Automatic segmentation facilitates quantitative analysis of fat metaplasia to improve diagnosis and prognosis of axial spondyloarthritis. Graphical Abstract

show abstract

Squeeze-and-Attention Networks for Semantic Segmentation

Cited by 4 publications

References 49 publications

Dynamic Region-Aware Convolution

Dynamic Region-Aware Convolution

Online Ensemble Model Compression Using Knowledge Distillation

Automatic segmentation of fat metaplasia on sacroiliac joint MRI using deep learning

Contact Info

Product

Resources

About