A sparsity-based stochastic pooling mechanism for deep convolutional neural networks

Song, Zhenhua; Liu, Yan; Song, Rong; Chen, Zhenguang; Yang, Jianyong; Zhang, Chao; Jiang, Qing

doi:10.1016/j.neunet.2018.05.015

Cited by 29 publications

(22 citation statements)

References 16 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…To a certain extent, it will alleviate the lack of medical service capacity in remote areas and the unbalanced development of medical standards [29]. Medical breakthroughs in geographical limits are conducive to the development of clinical research and improving the level of diagnosis and treatment of small and medium-sized physicians.…”

Section: Medical Internet Of Thingsmentioning

confidence: 99%

Head CT Image Convolution Feature Segmentation and Morphological Filtering for Densely Matching Points of IoTs

Yao

Xia

et al. 2020

IEEE Access

View full text Add to dashboard Cite

With the rapid application of medical imaging technology, the number of medical images is increasing and the form is gradually diversified. The management and retrieval of medical images is an urgent problem to solve. Traditional image matching technology has many limitations due to the complexity of its manual labeling. Therefore, content-based medical image retrieval is a new method to solve this problem. The technique uses the visual attributes contained in the image to establish the feature index of the image, adopts Sparse Connectivity and weight sharing of CNN (Convolutional Neural Network) to obtain the image feature, and then perform feature similarity matching between the query image and the image in the image database, and implement retrieval according to the matching result. The intensive matching based on the Internet of Things has great potential in the application of medical image 3D reconstruction. Compared with the traditional algorithm based on convolution feature segmentation and morphological filtering, the biggest advantage of the proposed algorithm is the computational efficiency, aiming at the same image. The efficiency of dense matching based on the Internet of Things has greatly improved. At the same time, the problem of occlusion, moving objects, rotation, and scaling in multi-view images is well handled. Experiments show that the proposed method can extract more dense and reliable matching points for multi-view images, which is beneficial to the subsequent 3D reconstruction.

show abstract

Section: Medical Internet Of Thingsmentioning

confidence: 99%

Head CT Image Convolution Feature Segmentation and Morphological Filtering for Densely Matching Points of IoTs

Yao

Xia

et al. 2020

IEEE Access

View full text Add to dashboard Cite

show abstract

“…The objective of the pooling layer is to achieve robustness to illumination changes and position variations with invariance to feature transformations [25]. In practice, most incipient CNNs have employed popular pooling methods, such as average pooling and maximum pooling [26]. Average pooling considers all the elements in the pooling regions to prevent variance increases while retaining the background information [26,27], whereas maximum pooling only captures the foreground texture information of the strongest activation as a representative feature of a region of an image, as shown in Figure 2 [24].…”

Section: Hybrid Poolingmentioning

confidence: 99%

“…In practice, most incipient CNNs have employed popular pooling methods, such as average pooling and maximum pooling [26]. Average pooling considers all the elements in the pooling regions to prevent variance increases while retaining the background information [26,27], whereas maximum pooling only captures the foreground texture information of the strongest activation as a representative feature of a region of an image, as shown in Figure 2 [24]. However, Boureau et al [28] noted that there are some drawbacks, as maximum and average pooling may lose information representing the background and foreground, respectively [26].…”

Section: Hybrid Poolingmentioning

confidence: 99%

“…Average pooling considers all the elements in the pooling regions to prevent variance increases while retaining the background information [26,27], whereas maximum pooling only captures the foreground texture information of the strongest activation as a representative feature of a region of an image, as shown in Figure 2 [24]. However, Boureau et al [28] noted that there are some drawbacks, as maximum and average pooling may lose information representing the background and foreground, respectively [26]. Thus, a concatenated pooling method combining the characteristics of maximum and average pooling was utilized in this study to prevent information loss near the representative features because of the spatial characteristics of concrete surface damage.…”

Section: Hybrid Poolingmentioning

confidence: 99%

See 1 more Smart Citation

Automatic Concrete Damage Recognition Using Multi-Level Attention Convolutional Neural Network

et al. 2020

View full text Add to dashboard Cite

There has been an increase in the deterioration of buildings and infrastructure in dense urban regions, and several defects in the structures are being exposed. To ensure the effective diagnosis of building conditions, vision-based automatic damage recognition techniques have been developed. However, conventional image processing techniques have some limitations in real-world situations owing to their manual feature extraction approach. To overcome these limitations, a convolutional neural network-based image recognition technique was adopted in this study, and a convolution-based concrete multi-damage recognition neural network (CMDnet) was developed. The image datasets consisted of 1981 types of concrete surface damages, including surface cracks, rebar exposure and delamination, as well as intact. Furthermore, it was experimentally demonstrated that the proposed model could accurately classify the damage types. The results obtained in this study reveal that the proposed model can recognize the different damage types from digital images of the surfaces of concrete structures. The trained CMDnet demonstrated a damage-detection accuracy of 98.9%. Moreover, the proposed model could be applied in automatic damage detection networks to achieve superior performance with regard to concrete surface damage detection and recognition, as well as accelerating efficient damage identification during the diagnosis of deteriorating structures used in civil engineering applications.

show abstract

“…Therefore, the learning network used in this paper is also a guided learning network. A convolutional network is essentially an input-to-output mapping that learns a large number of mappings between input and output without the need for any precise mathematical expression between input and output, as long as it is known [32]. The mode trains the convolutional network, and the network has the ability to map between input and output pairs.…”

Section: B Convolutional Network Training Processmentioning

confidence: 99%

Magnetic Resonance Image Segmentation Based on Multi-Scale Convolutional Neural Network

Hao¹,

Li²,

Hou

2020

IEEE Access

View full text Add to dashboard Cite

The incidence of nervous system and soft tissue is getting higher, and nuclear magnetic resonance is the preferred method of examination and is widely used. The segmentation of brain magnetic resonance (MR) images is the key to the subsequent operations such as three-dimensional reconstruction, quantitative analysis of normal tissues and diseased tissues. The accuracy of image segmentation is critical in the doctor's assessment with the location, shape, and size of the lesion tissue, as well as the determination of the disease and the correct diagnosis plan. The results of this study indicate that the multi-scale convolutional neural network (MSCNN) model can segment brain tumors accurately and effectively. Through multi-scale input, this paper overcomes the need to select specific input scales according to the size of the tumor, accommodate more neighborhood information from various angles, and adapt to different tumor sizes, thus improving the segmentation accuracy of brain tumors. Based on the same accuracy, the segmentation speed is accelerated to ensure the real-time segmentation further. This method can effectively segment the brain lesion tissue in the nuclear magnetic resonance image, which improves the generalization ability. It can be used for identifying the brain lesion tissue of the nuclear magnetic resonance medical image. INDEX TERMS Magnetic resonance (MR), subsequent operation, image segmentation, multi-scale convolutional neural network (MSCNN).

show abstract

A sparsity-based stochastic pooling mechanism for deep convolutional neural networks

Cited by 29 publications

References 16 publications

Head CT Image Convolution Feature Segmentation and Morphological Filtering for Densely Matching Points of IoTs

Head CT Image Convolution Feature Segmentation and Morphological Filtering for Densely Matching Points of IoTs

Automatic Concrete Damage Recognition Using Multi-Level Attention Convolutional Neural Network

Magnetic Resonance Image Segmentation Based on Multi-Scale Convolutional Neural Network

Contact Info

Product

Resources

About