A new spectral matching algorithm is proposed by us- ing nonsubsampled contourlet transform and scale-invariant fea- ture transform. The nonsubsampled contourlet transform is used to decompose an image into a low freq...A new spectral matching algorithm is proposed by us- ing nonsubsampled contourlet transform and scale-invariant fea- ture transform. The nonsubsampled contourlet transform is used to decompose an image into a low frequency image and several high frequency images, and the scale-invariant feature transform is employed to extract feature points from the low frequency im- age. A proximity matrix is constructed for the feature points of two related images. By singular value decomposition of the proximity matrix, a matching matrix （or matching result） reflecting the match- ing degree among feature points is obtained. Experimental results indicate that the proposed algorithm can reduce time complexity and possess a higher accuracy.展开更多
In this paper, we proposed a registration method by combining the morphological component analysis(MCA) and scale-invariant feature transform(SIFT) algorithm. This method uses the perception dictionaries,and combines ...In this paper, we proposed a registration method by combining the morphological component analysis(MCA) and scale-invariant feature transform(SIFT) algorithm. This method uses the perception dictionaries,and combines the Basis-Pursuit algorithm and the Total-Variation regularization scheme to extract the cartoon part containing basic geometrical information from the original image, and is stable and unsusceptible to noise interference. Then a smaller number of the distinctive key points will be obtained by using the SIFT algorithm based on the cartoon part of the original image. Matching the key points by the constrained Euclidean distance,we will obtain a more correct and robust matching result. The experimental results show that the geometrical transform parameters inferred by the matched key points based on MCA+SIFT registration method are more exact than the ones based on the direct SIFT algorithm.展开更多
To meet the needs in the fundus examination,including outlook widening,pathology tracking,etc.,this paper describes a robust feature-based method for fully-automatic mosaic of the curved human retinal images photograp...To meet the needs in the fundus examination,including outlook widening,pathology tracking,etc.,this paper describes a robust feature-based method for fully-automatic mosaic of the curved human retinal images photographed by a fundus microscope. The kernel of this new algorithm is the scale-,rotation-and illumination-invariant interest point detector & feature descriptor-Scale-Invariant Feature Transform. When matched interest points according to second-nearest-neighbor strategy,the parameters of the model are estimated using the correct matches of the interest points,extracted by a new inlier identification scheme based on Sampson distance from putative sets. In order to preserve image features,bilinear warping and multi-band blending techniques are used to create panoramic retinal images. Experiments show that the proposed method works well with rejection error in 0.3 pixels,even for those cases where the retinal images without discernable vascular structure in contrast to the state-of-the-art algorithms.展开更多
Content-based satellite image registration is a difficult issue in the fields of remote sensing and image processing. The difficulty is more significant in the case of matching multisource remote sensing images which ...Content-based satellite image registration is a difficult issue in the fields of remote sensing and image processing. The difficulty is more significant in the case of matching multisource remote sensing images which suffer from illumination, rotation, and source differences. The scale-invariant feature transform （SIFT） algorithm has been used successfully in satellite image registration problems. Also, many researchers have applied a local SIFT descriptor to improve the image retrieval process. Despite its robustness, this algorithm has some difficulties with the quality and quantity of the extracted local feature points in multisource remote sensing. Furthermore, high dimensionality of the local features extracted by SIFT results in time-consuming computational processes alongside high storage requirements for saving the relevant information, which are important factors in content-based image retrieval （CBIR） applications. In this paper, a novel method is introduced to transform the local SIFT features to global features for multisource remote sensing. The quality and quantity of SIFT local features have been enhanced by applying contrast equalization on images in a pre-processing stage. Considering the local features of each image in the reference database as a separate class, linear discriminant analysis （LDA） is used to transform the local features to global features while reducing di- mensionality of the feature space. This will also significantly reduce the computational time and storage required. Applying the trained kernel on verification data and mapping them showed a successful retrieval rate of 91.67% for test feature points.展开更多
Contraposing the need of the robust digital watermark for the copyright protection field, a new digital watermarking algorithm in the non-subsampled contourlet transform （NSCT） domain is proposed. The largest energy...Contraposing the need of the robust digital watermark for the copyright protection field, a new digital watermarking algorithm in the non-subsampled contourlet transform （NSCT） domain is proposed. The largest energy sub-band after NSCT is selected to embed watermark. The watermark is embedded into scaleinvariant feature transform （SIFT） regions. During embedding, the initial region is divided into some cirque sub-regions with the same area, and each watermark bit is embedded into one sub-region. Extensive simulation results and comparisons show that the algorithm gets a good trade-off of invisibility, robustness and capacity, thus obtaining good quality of the image while being able to effectively resist common image processing, and geometric and combo attacks, and normalized similarity is almost all reached.展开更多
Dear Editor,This letter proposes an end-to-end feature disentangled Transformer(FDTs)for entanglement-free and semantic feature representation to enable accurate and trustworthy pathology grading of squamous cell carc...Dear Editor,This letter proposes an end-to-end feature disentangled Transformer(FDTs)for entanglement-free and semantic feature representation to enable accurate and trustworthy pathology grading of squamous cell carcinoma(SCC).Existing vision transformers(ViTs)can implement representation learning for SCC grading,however,they all adopt the class-patch token fuzzy mapping for pattern prediction probability or window down-sampling to enhance the representation to contextual information.展开更多
Remote sensing plays a pivotal role in environmental monitoring,disaster relief,and urban planning,where accurate scene classification of aerial images is essential.However,conventional convolutional neural networks(C...Remote sensing plays a pivotal role in environmental monitoring,disaster relief,and urban planning,where accurate scene classification of aerial images is essential.However,conventional convolutional neural networks(CNNs)struggle with long-range dependencies and preserving high-resolution features,limiting their effectiveness in complex aerial image analysis.To address these challenges,we propose a Hybrid HRNet-Swin Transformer model that synergizes the strengths of HRNet-W48 for high-resolution segmentation and the Swin Transformer for global feature extraction.This hybrid architecture ensures robust multi-scale feature fusion,capturing fine-grained details and broader contextual relationships in aerial imagery.Our methodology begins with preprocessing steps,including normalization,histogram equalization,and noise reduction,to enhance input data quality.The HRNet-W48 backbone maintains high-resolution feature maps throughout the network,enabling precise segmentation,while the Swin Transformer leverages hierarchical self-attention to model long-range dependencies efficiently.By integrating these components,our model achieves superior performance in segmentation and classification tasks compared to traditional CNNs and standalone transformer models.We evaluate our approach on two benchmark datasets:UC Merced and WHU-RS19.Experimental results demonstrate that the proposed hybrid model outperforms existing methods,achieving state-of-the-art accuracy while maintaining computational efficiency.Specifically,it excels in preserving fine spatial details and contextual understanding,critical for applications like land-use classification and disaster assessment.展开更多
Generative image steganography is a technique that directly generates stego images from secret infor-mation.Unlike traditional methods,it theoretically resists steganalysis because there is no cover image.Currently,th...Generative image steganography is a technique that directly generates stego images from secret infor-mation.Unlike traditional methods,it theoretically resists steganalysis because there is no cover image.Currently,the existing generative image steganography methods generally have good steganography performance,but there is still potential room for enhancing both the quality of stego images and the accuracy of secret information extraction.Therefore,this paper proposes a generative image steganography algorithm based on attribute feature transformation and invertible mapping rule.Firstly,the reference image is disentangled by a content and an attribute encoder to obtain content features and attribute features,respectively.Then,a mean mapping rule is introduced to map the binary secret information into a noise vector,conforming to the distribution of attribute features.This noise vector is input into the generator to produce the attribute transformed stego image with the content feature of the reference image.Additionally,we design an adversarial loss,a reconstruction loss,and an image diversity loss to train the proposed model.Experimental results demonstrate that the stego images generated by the proposed method are of high quality,with an average extraction accuracy of 99.4%for the hidden information.Furthermore,since the stego image has a uniform distribution similar to the attribute-transformed image without secret information,it effectively resists both subjective and objective steganalysis.展开更多
Nonlinear transforms have significantly advanced learned image compression(LIC),particularly using residual blocks.This transform enhances the nonlinear expression ability and obtain compact feature representation by ...Nonlinear transforms have significantly advanced learned image compression(LIC),particularly using residual blocks.This transform enhances the nonlinear expression ability and obtain compact feature representation by enlarging the receptive field,which indicates how the convolution process extracts features in a high dimensional feature space.However,its functionality is restricted to the spatial dimension and network depth,limiting further improvements in network performance due to insufficient information interaction and representation.Crucially,the potential of high dimensional feature space in the channel dimension and the exploration of network width/resolution remain largely untapped.In this paper,we consider nonlinear transforms from the perspective of feature space,defining high-dimensional feature spaces in different dimensions and investigating the specific effects.Firstly,we introduce the dimension increasing and decreasing transforms in both channel and spatial dimensions to obtain high dimensional feature space and achieve better feature extraction.Secondly,we design a channel-spatial fusion residual transform(CSR),which incorporates multi-dimensional transforms for a more effective representation.Furthermore,we simplify the proposed fusion transform to obtain a slim architecture(CSR-sm),balancing network complexity and compression performance.Finally,we build the overall network with stacked CSR transforms to achieve better compression and reconstruction.Experimental results demonstrate that the proposed method can achieve superior ratedistortion performance compared to the existing LIC methods and traditional codecs.Specifically,our proposed method achieves 9.38%BD-rate reduction over VVC on Kodak dataset.展开更多
A method is proposed for the analysis of vibration signals from components ofrotating machines, based on the wavelet packet transformation (WPT) and the underlying physicalconcepts of modulation mechanism. The method ...A method is proposed for the analysis of vibration signals from components ofrotating machines, based on the wavelet packet transformation (WPT) and the underlying physicalconcepts of modulation mechanism. The method provides a finer analysis and better time-frequencylocalization capabilities than any other analysis methods. Both details and approximations are splitinto finer components and result in better-localized frequency ranges corresponding to each node ofa wavelet packet tree. For the punpose of feature extraction, a hard threshold is given and theenergy of the coefficients above the threshold is used, as a criterion for the selection of the bestvector. The feature extraction of a vibration signal is accomplished by computing thereconstruction signal and its spectrum. When applied to a rolling bear vibration signal featureextraction, the proposed method can lead to be very effective.展开更多
It is an important precondition for machine fault diagnosis that vibrationsignal can be extracted effectively. Based on the characteristic of noise interfused during thecourse of sampling vibration signal, independent...It is an important precondition for machine fault diagnosis that vibrationsignal can be extracted effectively. Based on the characteristic of noise interfused during thecourse of sampling vibration signal, independent component analysis (ICA) method is combined withwavelet to de-noise. Firstly, The sampled signal can be separated with ICA, then the function offrequency band chosen with multi-resolution wavelet transform can be used to judge whether thestochastic disturbance singular signal is interfused. By these ways, the vibration signals can beextracted effectively, which provides favorable condition for subsequent feature detection ofvibration signal and fault diagnosis.展开更多
Image registration is an indispensable component in multi-source remote sensing image processing. In this paper, we put forward a remote sensing image registration method by including an improved multi-scale and multi...Image registration is an indispensable component in multi-source remote sensing image processing. In this paper, we put forward a remote sensing image registration method by including an improved multi-scale and multi-direction Harris algorithm and a novel compound feature. Multi-scale circle Gaussian combined invariant moments and multi-direction gray level co-occurrence matrix are extracted as features for image matching. The proposed algorithm is evaluated on numerous multi-source remote sensor images with noise and illumination changes. Extensive experimental studies prove that our proposed method is capable of receiving stable and even distribution of key points as well as obtaining robust and accurate correspondence matches. It is a promising scheme in multi-source remote sensing image registration.展开更多
Recently, there have been some attempts of Transformer in 3D point cloud classification. In order to reduce computations, most existing methods focus on local spatial attention,but ignore their content and fail to est...Recently, there have been some attempts of Transformer in 3D point cloud classification. In order to reduce computations, most existing methods focus on local spatial attention,but ignore their content and fail to establish relationships between distant but relevant points. To overcome the limitation of local spatial attention, we propose a point content-based Transformer architecture, called PointConT for short. It exploits the locality of points in the feature space(content-based), which clusters the sampled points with similar features into the same class and computes the self-attention within each class, thus enabling an effective trade-off between capturing long-range dependencies and computational complexity. We further introduce an inception feature aggregator for point cloud classification, which uses parallel structures to aggregate high-frequency and low-frequency information in each branch separately. Extensive experiments show that our PointConT model achieves a remarkable performance on point cloud shape classification. Especially, our method exhibits 90.3% Top-1 accuracy on the hardest setting of ScanObjectN N. Source code of this paper is available at https://github.com/yahuiliu99/PointC onT.展开更多
基金supported by the National Natural Science Foundation of China (6117212711071002)+1 种基金the Specialized Research Fund for the Doctoral Program of Higher Education (20113401110006)the Innovative Research Team of 211 Project in Anhui University (KJTD007A)
文摘A new spectral matching algorithm is proposed by us- ing nonsubsampled contourlet transform and scale-invariant fea- ture transform. The nonsubsampled contourlet transform is used to decompose an image into a low frequency image and several high frequency images, and the scale-invariant feature transform is employed to extract feature points from the low frequency im- age. A proximity matrix is constructed for the feature points of two related images. By singular value decomposition of the proximity matrix, a matching matrix （or matching result） reflecting the match- ing degree among feature points is obtained. Experimental results indicate that the proposed algorithm can reduce time complexity and possess a higher accuracy.
基金the National Science Foundation of China(No.61471185)the Natural Science Foundation of Shandong Province(No.ZR2016FM21)+1 种基金Shandong Province Science and Technology Plan Project(No.2015GSF116001)Yantai City Key Research and Development Plan Project(Nos.2014ZH157 and2016ZH057)
文摘In this paper, we proposed a registration method by combining the morphological component analysis(MCA) and scale-invariant feature transform(SIFT) algorithm. This method uses the perception dictionaries,and combines the Basis-Pursuit algorithm and the Total-Variation regularization scheme to extract the cartoon part containing basic geometrical information from the original image, and is stable and unsusceptible to noise interference. Then a smaller number of the distinctive key points will be obtained by using the SIFT algorithm based on the cartoon part of the original image. Matching the key points by the constrained Euclidean distance,we will obtain a more correct and robust matching result. The experimental results show that the geometrical transform parameters inferred by the matched key points based on MCA+SIFT registration method are more exact than the ones based on the direct SIFT algorithm.
基金Program for NewCentury Excellent Talents in UniversityGrant number:50051+1 种基金The Key Project for Technology Research of Ministry Education of ChinaCrant number:106030
文摘To meet the needs in the fundus examination,including outlook widening,pathology tracking,etc.,this paper describes a robust feature-based method for fully-automatic mosaic of the curved human retinal images photographed by a fundus microscope. The kernel of this new algorithm is the scale-,rotation-and illumination-invariant interest point detector & feature descriptor-Scale-Invariant Feature Transform. When matched interest points according to second-nearest-neighbor strategy,the parameters of the model are estimated using the correct matches of the interest points,extracted by a new inlier identification scheme based on Sampson distance from putative sets. In order to preserve image features,bilinear warping and multi-band blending techniques are used to create panoramic retinal images. Experiments show that the proposed method works well with rejection error in 0.3 pixels,even for those cases where the retinal images without discernable vascular structure in contrast to the state-of-the-art algorithms.
文摘Content-based satellite image registration is a difficult issue in the fields of remote sensing and image processing. The difficulty is more significant in the case of matching multisource remote sensing images which suffer from illumination, rotation, and source differences. The scale-invariant feature transform （SIFT） algorithm has been used successfully in satellite image registration problems. Also, many researchers have applied a local SIFT descriptor to improve the image retrieval process. Despite its robustness, this algorithm has some difficulties with the quality and quantity of the extracted local feature points in multisource remote sensing. Furthermore, high dimensionality of the local features extracted by SIFT results in time-consuming computational processes alongside high storage requirements for saving the relevant information, which are important factors in content-based image retrieval （CBIR） applications. In this paper, a novel method is introduced to transform the local SIFT features to global features for multisource remote sensing. The quality and quantity of SIFT local features have been enhanced by applying contrast equalization on images in a pre-processing stage. Considering the local features of each image in the reference database as a separate class, linear discriminant analysis （LDA） is used to transform the local features to global features while reducing di- mensionality of the feature space. This will also significantly reduce the computational time and storage required. Applying the trained kernel on verification data and mapping them showed a successful retrieval rate of 91.67% for test feature points.
基金supported by the National Natural Science Foundation of China(61379010)the Natural Science Basic Research Plan in Shaanxi Province of China(2015JM6293)
文摘Contraposing the need of the robust digital watermark for the copyright protection field, a new digital watermarking algorithm in the non-subsampled contourlet transform （NSCT） domain is proposed. The largest energy sub-band after NSCT is selected to embed watermark. The watermark is embedded into scaleinvariant feature transform （SIFT） regions. During embedding, the initial region is divided into some cirque sub-regions with the same area, and each watermark bit is embedded into one sub-region. Extensive simulation results and comparisons show that the algorithm gets a good trade-off of invisibility, robustness and capacity, thus obtaining good quality of the image while being able to effectively resist common image processing, and geometric and combo attacks, and normalized similarity is almost all reached.
基金supported by the National Natural Science Foundation of China(62272078)the Chongqing Natural Science Foundation(CSTB2023NSCQ-LZX0069).
文摘Dear Editor,This letter proposes an end-to-end feature disentangled Transformer(FDTs)for entanglement-free and semantic feature representation to enable accurate and trustworthy pathology grading of squamous cell carcinoma(SCC).Existing vision transformers(ViTs)can implement representation learning for SCC grading,however,they all adopt the class-patch token fuzzy mapping for pattern prediction probability or window down-sampling to enhance the representation to contextual information.
基金supported by the ITP(Institute of Information&Communications Technology Planning&Evaluation)-ICAN(ICT Challenge and Advanced Network of HRD)(ITP-2025-RS-2022-00156326,33)grant funded by the Korea government(Ministry of Science and ICT)the Deanship of Research and Graduate Studies at King Khalid University for funding this work through the Large Group Project under grant number(RGP2/568/45)the Deanship of Scientific Research at Northern Border University,Arar,Saudi Arabia,for funding this research work through the Project Number"NBU-FFR-2025-231-03".
文摘Remote sensing plays a pivotal role in environmental monitoring,disaster relief,and urban planning,where accurate scene classification of aerial images is essential.However,conventional convolutional neural networks(CNNs)struggle with long-range dependencies and preserving high-resolution features,limiting their effectiveness in complex aerial image analysis.To address these challenges,we propose a Hybrid HRNet-Swin Transformer model that synergizes the strengths of HRNet-W48 for high-resolution segmentation and the Swin Transformer for global feature extraction.This hybrid architecture ensures robust multi-scale feature fusion,capturing fine-grained details and broader contextual relationships in aerial imagery.Our methodology begins with preprocessing steps,including normalization,histogram equalization,and noise reduction,to enhance input data quality.The HRNet-W48 backbone maintains high-resolution feature maps throughout the network,enabling precise segmentation,while the Swin Transformer leverages hierarchical self-attention to model long-range dependencies efficiently.By integrating these components,our model achieves superior performance in segmentation and classification tasks compared to traditional CNNs and standalone transformer models.We evaluate our approach on two benchmark datasets:UC Merced and WHU-RS19.Experimental results demonstrate that the proposed hybrid model outperforms existing methods,achieving state-of-the-art accuracy while maintaining computational efficiency.Specifically,it excels in preserving fine spatial details and contextual understanding,critical for applications like land-use classification and disaster assessment.
基金supported in part by the National Natural Science Foundation of China(Nos.62202234,62401270)the China Postdoctoral Science Foundation(No.2023M741778)the Natural Science Foundation of Jiangsu Province(Nos.BK20240706,BK20240694).
文摘Generative image steganography is a technique that directly generates stego images from secret infor-mation.Unlike traditional methods,it theoretically resists steganalysis because there is no cover image.Currently,the existing generative image steganography methods generally have good steganography performance,but there is still potential room for enhancing both the quality of stego images and the accuracy of secret information extraction.Therefore,this paper proposes a generative image steganography algorithm based on attribute feature transformation and invertible mapping rule.Firstly,the reference image is disentangled by a content and an attribute encoder to obtain content features and attribute features,respectively.Then,a mean mapping rule is introduced to map the binary secret information into a noise vector,conforming to the distribution of attribute features.This noise vector is input into the generator to produce the attribute transformed stego image with the content feature of the reference image.Additionally,we design an adversarial loss,a reconstruction loss,and an image diversity loss to train the proposed model.Experimental results demonstrate that the stego images generated by the proposed method are of high quality,with an average extraction accuracy of 99.4%for the hidden information.Furthermore,since the stego image has a uniform distribution similar to the attribute-transformed image without secret information,it effectively resists both subjective and objective steganalysis.
基金supported by the Key Program of the National Natural Science Foundation of China(Grant No.62031013)Guangdong Province Key Construction Discipline Scientific Research Capacity Improvement Project(Grant No.2022ZDJS117).
文摘Nonlinear transforms have significantly advanced learned image compression(LIC),particularly using residual blocks.This transform enhances the nonlinear expression ability and obtain compact feature representation by enlarging the receptive field,which indicates how the convolution process extracts features in a high dimensional feature space.However,its functionality is restricted to the spatial dimension and network depth,limiting further improvements in network performance due to insufficient information interaction and representation.Crucially,the potential of high dimensional feature space in the channel dimension and the exploration of network width/resolution remain largely untapped.In this paper,we consider nonlinear transforms from the perspective of feature space,defining high-dimensional feature spaces in different dimensions and investigating the specific effects.Firstly,we introduce the dimension increasing and decreasing transforms in both channel and spatial dimensions to obtain high dimensional feature space and achieve better feature extraction.Secondly,we design a channel-spatial fusion residual transform(CSR),which incorporates multi-dimensional transforms for a more effective representation.Furthermore,we simplify the proposed fusion transform to obtain a slim architecture(CSR-sm),balancing network complexity and compression performance.Finally,we build the overall network with stacked CSR transforms to achieve better compression and reconstruction.Experimental results demonstrate that the proposed method can achieve superior ratedistortion performance compared to the existing LIC methods and traditional codecs.Specifically,our proposed method achieves 9.38%BD-rate reduction over VVC on Kodak dataset.
文摘A method is proposed for the analysis of vibration signals from components ofrotating machines, based on the wavelet packet transformation (WPT) and the underlying physicalconcepts of modulation mechanism. The method provides a finer analysis and better time-frequencylocalization capabilities than any other analysis methods. Both details and approximations are splitinto finer components and result in better-localized frequency ranges corresponding to each node ofa wavelet packet tree. For the punpose of feature extraction, a hard threshold is given and theenergy of the coefficients above the threshold is used, as a criterion for the selection of the bestvector. The feature extraction of a vibration signal is accomplished by computing thereconstruction signal and its spectrum. When applied to a rolling bear vibration signal featureextraction, the proposed method can lead to be very effective.
基金This project is supported by National Natural Science Foundation of China (No.50275154) Municipal Natural Science Foundation of Chongqing, China (No.8773).
文摘It is an important precondition for machine fault diagnosis that vibrationsignal can be extracted effectively. Based on the characteristic of noise interfused during thecourse of sampling vibration signal, independent component analysis (ICA) method is combined withwavelet to de-noise. Firstly, The sampled signal can be separated with ICA, then the function offrequency band chosen with multi-resolution wavelet transform can be used to judge whether thestochastic disturbance singular signal is interfused. By these ways, the vibration signals can beextracted effectively, which provides favorable condition for subsequent feature detection ofvibration signal and fault diagnosis.
基金supported by National Nature Science Foundation of China (Nos. 61462046 and 61762052)Natural Science Foundation of Jiangxi Province (Nos. 20161BAB202049 and 20161BAB204172)+2 种基金the Bidding Project of the Key Laboratory of Watershed Ecology and Geographical Environment Monitoring, NASG (Nos. WE2016003, WE2016013 and WE2016015)the Science and Technology Research Projects of Jiangxi Province Education Department (Nos. GJJ160741, GJJ170632 and GJJ170633)the Art Planning Project of Jiangxi Province (Nos. YG2016250 and YG2017381)
文摘Image registration is an indispensable component in multi-source remote sensing image processing. In this paper, we put forward a remote sensing image registration method by including an improved multi-scale and multi-direction Harris algorithm and a novel compound feature. Multi-scale circle Gaussian combined invariant moments and multi-direction gray level co-occurrence matrix are extracted as features for image matching. The proposed algorithm is evaluated on numerous multi-source remote sensor images with noise and illumination changes. Extensive experimental studies prove that our proposed method is capable of receiving stable and even distribution of key points as well as obtaining robust and accurate correspondence matches. It is a promising scheme in multi-source remote sensing image registration.
基金supported in part by the Nationa Natural Science Foundation of China (61876011)the National Key Research and Development Program of China (2022YFB4703700)+1 种基金the Key Research and Development Program 2020 of Guangzhou (202007050002)the Key-Area Research and Development Program of Guangdong Province (2020B090921003)。
文摘Recently, there have been some attempts of Transformer in 3D point cloud classification. In order to reduce computations, most existing methods focus on local spatial attention,but ignore their content and fail to establish relationships between distant but relevant points. To overcome the limitation of local spatial attention, we propose a point content-based Transformer architecture, called PointConT for short. It exploits the locality of points in the feature space(content-based), which clusters the sampled points with similar features into the same class and computes the self-attention within each class, thus enabling an effective trade-off between capturing long-range dependencies and computational complexity. We further introduce an inception feature aggregator for point cloud classification, which uses parallel structures to aggregate high-frequency and low-frequency information in each branch separately. Extensive experiments show that our PointConT model achieves a remarkable performance on point cloud shape classification. Especially, our method exhibits 90.3% Top-1 accuracy on the hardest setting of ScanObjectN N. Source code of this paper is available at https://github.com/yahuiliu99/PointC onT.
