Small-Object Sensitive Segmentation Using Across Feature Map Attention

Sang, Shengtian; Zhou, Yuyin; Islam, Md Tauhidul; Xing, Lei

doi:10.1109/tpami.2022.3211171

Cited by 20 publications

(6 citation statements)

References 50 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Despite the remarkable progress in semantic segmentation, many methods still need help to meet the demands of lightweight processing. In recent years, some researchers have proposed a series of new approaches ( He et al, 2021 ; Sang et al, 2022 ), which adopt the design principles of network lightweight, aiming to bridge the gap between performance and efficiency. The emergence of these lightweight methods has provided robust solutions for lightweight semantic segmentation, further driving the advancements in this field.…”

Section: Related Workmentioning

confidence: 99%

Lightweight semantic segmentation network with configurable context and small object attention

Zhang,

Xu,

et al. 2023

Front. Comput. Neurosci.

View full text Add to dashboard Cite

The current semantic segmentation algorithms suffer from encoding feature distortion and small object feature loss. Context information exchange can effectively address the feature distortion problem, but it has the issue of fixed spatial range. Maintaining the input feature resolution can reduce the loss of small object information but would slow down the network’s operation speed. To tackle these problems, we propose a lightweight semantic segmentation network with configurable context and small object attention (CCSONet). CCSONet includes a long-short distance configurable context feature enhancement module (LSCFEM) and a small object attention decoding module (SOADM). The LSCFEM differs from the regular context exchange module by configuring long and short-range relevant features for the current feature, providing a broader and more flexible spatial range. The SOADM enhances the features of small objects by establishing correlations among objects of the same category, avoiding the introduction of redundancy issues caused by high-resolution features. On the Cityscapes and Camvid datasets, our network achieves the accuracy of 76.9 mIoU and 73.1 mIoU, respectively, while maintaining speeds of 87 FPS and 138 FPS. It outperforms other lightweight semantic segmentation algorithms in terms of accuracy.

show abstract

Section: Related Workmentioning

confidence: 99%

Lightweight semantic segmentation network with configurable context and small object attention

Zhang,

Xu,

et al. 2023

Front. Comput. Neurosci.

View full text Add to dashboard Cite

show abstract

“…Additionally, Y. Chong et al [35] introduced a contextaggregated edge network to merge context and edge information. S. Sang et al [36] enhance small object segmentation by leveraging feature correlations between small and large objects. Despite their merits, these methods still have limitations, such as being confined to natural images, not performing multi-class segmentation, or relying on specific objects present in the images.…”

Section: Related Workmentioning

confidence: 99%

YOSEG Enhancing Instance Segmentation of Medical Small Targets

Xia,

Zhu,

Mou

et al. 2023

Preprint

View full text Add to dashboard Cite

Image segmentation stands as a pivotal pillar in the vast realm of computer vision. Nonetheless, formidable challenges endure, particularly when confronted with tasks necessitating precision in delineating diminutive objects, as evidenced in the context of medical imaging for minute targets. Existing algorithms primarily cater to larger or medium-sized objects with specific dimensions or proportions. The persistent conundrum arises from the trifecta of minuscule target dimensions, feeble distinguishing features, and resultant subpar segmentation performance for diminutive objects. In response to these hurdles, our study introduces a pioneering solution christened You only look one segment network (YOSEG), adeptly crafted to tackle these issues head-on. Our comprehensive investigations involve two datasets: a proprietary collection comprising CT intracranial aneurysm data from multiple medical centers and the publicly available CBIS-DDSM dataset, which stands for Curated Breast Imaging Subset of DDSM and contains breast mass data. The outcomes of our approach substantiate remarkable strides in the precision of small object segmentation, marking a significant advancement in this critical field.

show abstract

“…Zhu et al [31] introduced an asymmetric Non-Local module, enabling attention computations by sampling representative points from key and value branches. Sang et al [32] proposed the AFMA attention mechanism module, quantifying the internal relationships between small and large objects of the same category in different-level features of original images, compensating for the loss of high-level feature information for small objects, and enhancing the segmentation of small targets.…”

Section: Related Workmentioning

confidence: 99%

Coal-Rock Crack Image Segmentation Based on Attention Mechanism and Multi-Scale Features

Jia,

Chen,

Zhang

et al. 2023

Preprint

View full text Add to dashboard Cite

Global contextual information and local information are crucial for the segmentation of coal rock fractures. For global information, most existing schemes directly input the feature maps output by CNN into Transformer for extraction. Due to the limitation of feature map size, Transformer is unable to further extract multi-scale information of cracks, making it unsuitable for cracks of different sizes; Moreover, the feature map contains data with low correlation with cracks, and directly inputting it into Transformer will affect the extraction of global information about cracks.Unet based networks are one of the mainstream solutions for extracting local detail information in images, but the low-level features of their skip connections contain redundant information, which directly leads to false positive segmentation results.Based on the above issues, this article designs the Channel Attention Atrus Spatial Pyramid Pooling module (TD-ASPP) and the Cross Attention Non-Local Block (CANB), and integrates them into Transunet to propose a coal rock fracture image segmentation based on attention mechanism and multi-scale features (MS-Unet).Firstly, the TD-ASPP module is used to capture multi-scale features of cracks, while strengthening channels with strong correlation with cracks and diluting channels with weak correlation, in order to further improve the accuracy of the transformer in calculating global information of cracks. Secondly, the CANB module can utilize advanced features to supervise and strengthen low-level features, further enhancing local details of cracks and suppressing false positives.

show abstract

Small-Object Sensitive Segmentation Using Across Feature Map Attention

Cited by 20 publications

References 50 publications

Lightweight semantic segmentation network with configurable context and small object attention

Lightweight semantic segmentation network with configurable context and small object attention

YOSEG Enhancing Instance Segmentation of Medical Small Targets

Coal-Rock Crack Image Segmentation Based on Attention Mechanism and Multi-Scale Features

Contact Info

Product

Resources

About