This study presents a novel and highly efficient superpixel algorithm, namely, depth-fused adaptive superpixel(DFASP), which can generate accurate superpixels in a degraded image. In many applications, particularly in...This study presents a novel and highly efficient superpixel algorithm, namely, depth-fused adaptive superpixel(DFASP), which can generate accurate superpixels in a degraded image. In many applications, particularly in actual scenes, vision degradation, such as motion blur, overexposure, and underexposure, often occurs. Well-known color-based superpixel algorithms are incapable of producing accurate superpixels in degraded images because of the ambiguity of color information caused by vision degradation. To eliminate this ambiguity, we use depth and color information to generate superpixels. We map the depth and color information to a high-dimensional feature space. Then, we develop a fast multilevel clustering algorithm to produce superpixels. Furthermore, we design an adaptive mechanism to adjust the color and depth information automatically during pixel clustering. Experimental results demonstrate that regardless of boundary recall, under segmentation error, run time, or achievable segmentation accuracy, DFASP is better than state-of-the-art superpixel methods.展开更多
Image captioning has gained increasing attention in recent years.Visual characteristics found in input images play a crucial role in generating high-quality captions.Prior studies have used visual attention mechanisms...Image captioning has gained increasing attention in recent years.Visual characteristics found in input images play a crucial role in generating high-quality captions.Prior studies have used visual attention mechanisms to dynamically focus on localized regions of the input image,improving the effectiveness of identifying relevant image regions at each step of caption generation.However,providing image captioning models with the capability of selecting the most relevant visual features from the input image and attending to them can significantly improve the utilization of these features.Consequently,this leads to enhanced captioning network performance.In light of this,we present an image captioning framework that efficiently exploits the extracted representations of the image.Our framework comprises three key components:the Visual Feature Detector module(VFD),the Visual Feature Visual Attention module(VFVA),and the language model.The VFD module is responsible for detecting a subset of the most pertinent features from the local visual features,creating an updated visual features matrix.Subsequently,the VFVA directs its attention to the visual features matrix generated by the VFD,resulting in an updated context vector employed by the language model to generate an informative description.Integrating the VFD and VFVA modules introduces an additional layer of processing for the visual features,thereby contributing to enhancing the image captioning model’s performance.Using the MS-COCO dataset,our experiments show that the proposed framework competes well with state-of-the-art methods,effectively leveraging visual representations to improve performance.The implementation code can be found here:https://github.com/althobhani/VFDICM(accessed on 30 July 2024).展开更多
The task of indoor visual localization, utilizing camera visual information for user pose calculation, was a core component of Augmented Reality (AR) and Simultaneous Localization and Mapping (SLAM). Existing indoor l...The task of indoor visual localization, utilizing camera visual information for user pose calculation, was a core component of Augmented Reality (AR) and Simultaneous Localization and Mapping (SLAM). Existing indoor localization technologies generally used scene-specific 3D representations or were trained on specific datasets, making it challenging to balance accuracy and cost when applied to new scenes. Addressing this issue, this paper proposed a universal indoor visual localization method based on efficient image retrieval. Initially, a Multi-Layer Perceptron (MLP) was employed to aggregate features from intermediate layers of a convolutional neural network, obtaining a global representation of the image. This approach ensured accurate and rapid retrieval of reference images. Subsequently, a new mechanism using Random Sample Consensus (RANSAC) was designed to resolve relative pose ambiguity caused by the essential matrix decomposition based on the five-point method. Finally, the absolute pose of the queried user image was computed, thereby achieving indoor user pose estimation. The proposed indoor localization method was characterized by its simplicity, flexibility, and excellent cross-scene generalization. Experimental results demonstrated a positioning error of 0.09 m and 2.14° on the 7Scenes dataset, and 0.15 m and 6.37° on the 12Scenes dataset. These results convincingly illustrated the outstanding performance of the proposed indoor localization method.展开更多
Objective image quality assessment(IQA)plays an important role in various visual communication systems,which can automatically and efficiently predict the perceived quality of images.The human eye is the ultimate eval...Objective image quality assessment(IQA)plays an important role in various visual communication systems,which can automatically and efficiently predict the perceived quality of images.The human eye is the ultimate evaluator for visual experience,thus the modeling of human visual system(HVS)is a core issue for objective IQA and visual experience optimization.The traditional model based on black box fitting has low interpretability and it is difficult to guide the experience optimization effectively,while the model based on physiological simulation is hard to integrate into practical visual communication services due to its high computational complexity.For bridging the gap between signal distortion and visual experience,in this paper,we propose a novel perceptual no-reference(NR)IQA algorithm based on structural computational modeling of HVS.According to the mechanism of the human brain,we divide the visual signal processing into a low-level visual layer,a middle-level visual layer and a high-level visual layer,which conduct pixel information processing,primitive information processing and global image information processing,respectively.The natural scene statistics(NSS)based features,deep features and free-energy based features are extracted from these three layers.The support vector regression(SVR)is employed to aggregate features to the final quality prediction.Extensive experimental comparisons on three widely used benchmark IQA databases(LIVE,CSIQ and TID2013)demonstrate that our proposed metric is highly competitive with or outperforms the state-of-the-art NR IQA measures.展开更多
It is illegal to spread and transmit pornographic images over internet,either in real or in artificial format.The traditional methods are designed to identify real pornographic images and they are less efficient in de...It is illegal to spread and transmit pornographic images over internet,either in real or in artificial format.The traditional methods are designed to identify real pornographic images and they are less efficient in dealing with artificial images.Therefore,criminals turn to release artificial pornographic images in some specific scenes,e.g.,in social networks.To efficiently identify artificial pornographic images,a novel bag-of-visual-words based approach is proposed in the work.In the bag-of-words(Bo W)framework,speeded-up robust feature(SURF)is adopted for feature extraction at first,then a visual vocabulary is constructed through K-means clustering and images are represented by an improved Bo W encoding method,and finally the visual words are fed into a learning machine for training and classification.Different from the traditional BoW method,the proposed method sets a weight on each visual word according to the number of features that each cluster contains.Moreover,a non-binary encoding method and cross-matching strategy are utilized to improve the discriminative power of the visual words.Experimental results indicate that the proposed method outperforms the traditional method.展开更多
In last few years,guided image fusion algorithms become more and more popular.However,the current algorithms cannot solve the halo artifacts.We propose an image fusion algorithm based on fast weighted guided filter.Fi...In last few years,guided image fusion algorithms become more and more popular.However,the current algorithms cannot solve the halo artifacts.We propose an image fusion algorithm based on fast weighted guided filter.Firstly,the source images are separated into a series of high and low frequency components.Secondly,three visual features of the source image are extracted to construct a decision graph model.Thirdly,a fast weighted guided filter is raised to optimize the result obtained in the previous step and reduce the time complexity by considering the correlation among neighboring pixels.Finally,the image obtained in the previous step is combined with the weight map to realize the image fusion.The proposed algorithm is applied to multi-focus,visible-infrared and multi-modal image respectively and the final results show that the algorithm effectively solves the halo artifacts of the merged images with higher efficiency,and is better than the traditional method considering subjective visual consequent and objective evaluation.展开更多
In medical research and clinical diagnosis, automated or computer-assisted classification and retrieval methods are highly desirable to offset the high cost of manual classification and manipulation by medical experts...In medical research and clinical diagnosis, automated or computer-assisted classification and retrieval methods are highly desirable to offset the high cost of manual classification and manipulation by medical experts. To facilitate the decision-making in the health-care and the related areas, in this paper, a two-step content-based medical image retrieval algorithm is proposed. Firstly, in the preprocessing step, the image segmentation is performed to distinguish image objects, and on the basis of the ...展开更多
Classifying the visual features in images to retrieve a specific image is a significant problem within the computer vision field especially when dealing with historical faded colored images.Thus,there were lots of eff...Classifying the visual features in images to retrieve a specific image is a significant problem within the computer vision field especially when dealing with historical faded colored images.Thus,there were lots of efforts trying to automate the classification operation and retrieve similar images accurately.To reach this goal,we developed a VGG19 deep convolutional neural network to extract the visual features from the images automatically.Then,the distances among the extracted features vectors are measured and a similarity score is generated using a Siamese deep neural network.The Siamese model built and trained at first from scratch but,it didn’t generated high evaluation metrices.Thus,we re-built it from VGG19 pre-trained deep learning model to generate higher evaluation metrices.Afterward,three different distance metrics combined with the Sigmoid activation function are experimented looking for the most accurate method formeasuring the similarities among the retrieved images.Reaching that the highest evaluation parameters generated using the Cosine distance metric.Moreover,the Graphics Processing Unit(GPU)utilized to run the code instead of running it on the Central Processing Unit(CPU).This step optimized the execution further since it expedited both the training and the retrieval time efficiently.After extensive experimentation,we reached satisfactory solution recording 0.98 and 0.99 F-score for the classification and for the retrieval,respectively.展开更多
In fMRI experiments on object representation in visual cortex, we designed two types of stimuli: one is the gray face image and its line drawing, and the other is the illusion and its corresponding completed illusion....In fMRI experiments on object representation in visual cortex, we designed two types of stimuli: one is the gray face image and its line drawing, and the other is the illusion and its corresponding completed illusion. Both of them have the same global features with different minute details so that the results of fMRI experiments can be compared with each other. The first kind of visual stimuli was used in a block design fMRI experiment, and the second was used in an event-related fMRI experiment. Comparing and analyzing interesting visual cortex activity patterns and blood oxygenation level dependent (BOLD)- fMRI signal, we obtained results to show some invariance of global features of visual images. A plau- sible explanation about the invariant mechanism is related with the cooperation of synchronized re- sponse to the global features of the visual image with a feedback of shape perception from higher cortex to cortex V1, namely the integration of global features and embodiment of sparse representation and distributed population code.展开更多
It is not clear whether the method used in functional brain-network related research can be applied to explore the feature binding mechanism of visual perception. In this study, we inves-tigated feature binding of col...It is not clear whether the method used in functional brain-network related research can be applied to explore the feature binding mechanism of visual perception. In this study, we inves-tigated feature binding of color and shape in visual perception. Functional magnetic resonance imaging data were collected from 38 healthy volunteers at rest and while performing a visual perception task to construct brain networks active during resting and task states. Results showed that brain regions involved in visual information processing were obviously activated during the task. The components were partitioned using a greedy algorithm, indicating the visual network existed during the resting state.Z-values in the vision-related brain regions were calculated, conifrming the dynamic balance of the brain network. Connectivity between brain regions was determined, and the result showed that occipital and lingual gyri were stable brain regions in the visual system network, the parietal lobe played a very important role in the binding process of color features and shape features, and the fusiform and inferior temporal gyri were crucial for processing color and shape information. Experimental ifndings indicate that understanding visual feature binding and cognitive processes will help establish computational models of vision, improve image recognition technology, and provide a new theoretical mechanism for feature binding in visual perception.展开更多
Existing glass segmentation networks have high computational complexity and large memory occupation,leading to high hardware requirements and time overheads for model inference,which is not conducive to efficiency-see...Existing glass segmentation networks have high computational complexity and large memory occupation,leading to high hardware requirements and time overheads for model inference,which is not conducive to efficiency-seeking real-time tasks such as autonomous driving.The inefficiency of the models is mainly due to employing homogeneous modules to process features of different layers.These modules require computationally intensive convolutions and weight calculation branches with numerous parameters to accommodate the differences in information across layers.We propose an efficient glass segmentation network(EGSNet)based on multi-level heterogeneous architecture and boundary awareness to balance the model performance and efficiency.EGSNet divides the feature layers from different stages into low-level understanding,semantic-level understanding,and global understanding with boundary guidance.Based on the information differences among the different layers,we further propose the multi-angle collaborative enhancement(MCE)module,which extracts the detailed information from shallow features,and the large-scale contextual feature extraction(LCFE)module to understand semantic logic through deep features.The models are trained and evaluated on the glass segmentation datasets HSO(Home-Scene-Oriented)and Trans10k-stuff,respectively,and EGSNet achieves the best efficiency and performance compared to advanced methods.In the HSO test set results,the IoU,Fβ,MAE(Mean Absolute Error),and BER(Balance Error Rate)of EGSNet are 0.804,0.847,0.084,and 0.085,and the GFLOPs(Giga Floating Point Operations Per Second)are only 27.15.Experimental results show that EGSNet significantly improves the efficiency of the glass segmentation task with better performance.展开更多
Visual question answering(VQA)has attracted more and more attention in computer vision and natural language processing.Scholars are committed to studying how to better integrate image features and text features to ach...Visual question answering(VQA)has attracted more and more attention in computer vision and natural language processing.Scholars are committed to studying how to better integrate image features and text features to achieve better results in VQA tasks.Analysis of all features may cause information redundancy and heavy computational burden.Attention mechanism is a wise way to solve this problem.However,using single attention mechanism may cause incomplete concern of features.This paper improves the attention mechanism method and proposes a hybrid attention mechanism that combines the spatial attention mechanism method and the channel attention mechanism method.In the case that the attention mechanism will cause the loss of the original features,a small portion of image features were added as compensation.For the attention mechanism of text features,a selfattention mechanism was introduced,and the internal structural features of sentences were strengthened to improve the overall model.The results show that attention mechanism and feature compensation add 6.1%accuracy to multimodal low-rank bilinear pooling network.展开更多
基金supported by the Science Technology Department of Zhejiang Province(No.LGG19F020010)the Department of Education of Zhejiang Province(No.Y201329938)the National Natural Science Foundation of China(No.61876167)
文摘This study presents a novel and highly efficient superpixel algorithm, namely, depth-fused adaptive superpixel(DFASP), which can generate accurate superpixels in a degraded image. In many applications, particularly in actual scenes, vision degradation, such as motion blur, overexposure, and underexposure, often occurs. Well-known color-based superpixel algorithms are incapable of producing accurate superpixels in degraded images because of the ambiguity of color information caused by vision degradation. To eliminate this ambiguity, we use depth and color information to generate superpixels. We map the depth and color information to a high-dimensional feature space. Then, we develop a fast multilevel clustering algorithm to produce superpixels. Furthermore, we design an adaptive mechanism to adjust the color and depth information automatically during pixel clustering. Experimental results demonstrate that regardless of boundary recall, under segmentation error, run time, or achievable segmentation accuracy, DFASP is better than state-of-the-art superpixel methods.
基金supported by the National Natural Science Foundation of China(Nos.U22A2034,62177047)High Caliber Foreign Experts Introduction Plan funded by MOST,and Central South University Research Programme of Advanced Interdisciplinary Studies(No.2023QYJC020).
文摘Image captioning has gained increasing attention in recent years.Visual characteristics found in input images play a crucial role in generating high-quality captions.Prior studies have used visual attention mechanisms to dynamically focus on localized regions of the input image,improving the effectiveness of identifying relevant image regions at each step of caption generation.However,providing image captioning models with the capability of selecting the most relevant visual features from the input image and attending to them can significantly improve the utilization of these features.Consequently,this leads to enhanced captioning network performance.In light of this,we present an image captioning framework that efficiently exploits the extracted representations of the image.Our framework comprises three key components:the Visual Feature Detector module(VFD),the Visual Feature Visual Attention module(VFVA),and the language model.The VFD module is responsible for detecting a subset of the most pertinent features from the local visual features,creating an updated visual features matrix.Subsequently,the VFVA directs its attention to the visual features matrix generated by the VFD,resulting in an updated context vector employed by the language model to generate an informative description.Integrating the VFD and VFVA modules introduces an additional layer of processing for the visual features,thereby contributing to enhancing the image captioning model’s performance.Using the MS-COCO dataset,our experiments show that the proposed framework competes well with state-of-the-art methods,effectively leveraging visual representations to improve performance.The implementation code can be found here:https://github.com/althobhani/VFDICM(accessed on 30 July 2024).
文摘The task of indoor visual localization, utilizing camera visual information for user pose calculation, was a core component of Augmented Reality (AR) and Simultaneous Localization and Mapping (SLAM). Existing indoor localization technologies generally used scene-specific 3D representations or were trained on specific datasets, making it challenging to balance accuracy and cost when applied to new scenes. Addressing this issue, this paper proposed a universal indoor visual localization method based on efficient image retrieval. Initially, a Multi-Layer Perceptron (MLP) was employed to aggregate features from intermediate layers of a convolutional neural network, obtaining a global representation of the image. This approach ensured accurate and rapid retrieval of reference images. Subsequently, a new mechanism using Random Sample Consensus (RANSAC) was designed to resolve relative pose ambiguity caused by the essential matrix decomposition based on the five-point method. Finally, the absolute pose of the queried user image was computed, thereby achieving indoor user pose estimation. The proposed indoor localization method was characterized by its simplicity, flexibility, and excellent cross-scene generalization. Experimental results demonstrated a positioning error of 0.09 m and 2.14° on the 7Scenes dataset, and 0.15 m and 6.37° on the 12Scenes dataset. These results convincingly illustrated the outstanding performance of the proposed indoor localization method.
基金This work was supported by National Natural Science Foundation of China(Nos.61831015 and 61901260)Key Research and Development Program of China(No.2019YFB1405902).
文摘Objective image quality assessment(IQA)plays an important role in various visual communication systems,which can automatically and efficiently predict the perceived quality of images.The human eye is the ultimate evaluator for visual experience,thus the modeling of human visual system(HVS)is a core issue for objective IQA and visual experience optimization.The traditional model based on black box fitting has low interpretability and it is difficult to guide the experience optimization effectively,while the model based on physiological simulation is hard to integrate into practical visual communication services due to its high computational complexity.For bridging the gap between signal distortion and visual experience,in this paper,we propose a novel perceptual no-reference(NR)IQA algorithm based on structural computational modeling of HVS.According to the mechanism of the human brain,we divide the visual signal processing into a low-level visual layer,a middle-level visual layer and a high-level visual layer,which conduct pixel information processing,primitive information processing and global image information processing,respectively.The natural scene statistics(NSS)based features,deep features and free-energy based features are extracted from these three layers.The support vector regression(SVR)is employed to aggregate features to the final quality prediction.Extensive experimental comparisons on three widely used benchmark IQA databases(LIVE,CSIQ and TID2013)demonstrate that our proposed metric is highly competitive with or outperforms the state-of-the-art NR IQA measures.
基金Projects(41001260,61173122,61573380) supported by the National Natural Science Foundation of ChinaProject(11JJ5044) supported by the Hunan Provincial Natural Science Foundation of China
文摘It is illegal to spread and transmit pornographic images over internet,either in real or in artificial format.The traditional methods are designed to identify real pornographic images and they are less efficient in dealing with artificial images.Therefore,criminals turn to release artificial pornographic images in some specific scenes,e.g.,in social networks.To efficiently identify artificial pornographic images,a novel bag-of-visual-words based approach is proposed in the work.In the bag-of-words(Bo W)framework,speeded-up robust feature(SURF)is adopted for feature extraction at first,then a visual vocabulary is constructed through K-means clustering and images are represented by an improved Bo W encoding method,and finally the visual words are fed into a learning machine for training and classification.Different from the traditional BoW method,the proposed method sets a weight on each visual word according to the number of features that each cluster contains.Moreover,a non-binary encoding method and cross-matching strategy are utilized to improve the discriminative power of the visual words.Experimental results indicate that the proposed method outperforms the traditional method.
基金supported by the National Natural Science Foundation of China(61472324 61671383)+1 种基金Shaanxi Key Industry Innovation Chain Project(2018ZDCXL-G-12-2 2019ZDLGY14-02-02)
文摘In last few years,guided image fusion algorithms become more and more popular.However,the current algorithms cannot solve the halo artifacts.We propose an image fusion algorithm based on fast weighted guided filter.Firstly,the source images are separated into a series of high and low frequency components.Secondly,three visual features of the source image are extracted to construct a decision graph model.Thirdly,a fast weighted guided filter is raised to optimize the result obtained in the previous step and reduce the time complexity by considering the correlation among neighboring pixels.Finally,the image obtained in the previous step is combined with the weight map to realize the image fusion.The proposed algorithm is applied to multi-focus,visible-infrared and multi-modal image respectively and the final results show that the algorithm effectively solves the halo artifacts of the merged images with higher efficiency,and is better than the traditional method considering subjective visual consequent and objective evaluation.
文摘In medical research and clinical diagnosis, automated or computer-assisted classification and retrieval methods are highly desirable to offset the high cost of manual classification and manipulation by medical experts. To facilitate the decision-making in the health-care and the related areas, in this paper, a two-step content-based medical image retrieval algorithm is proposed. Firstly, in the preprocessing step, the image segmentation is performed to distinguish image objects, and on the basis of the ...
基金The authors would like to thank the Deanship of Scientific Research at Umm Al-Qura University for supporting this work by Grant Code:(22UQU4400271DSR01).
文摘Classifying the visual features in images to retrieve a specific image is a significant problem within the computer vision field especially when dealing with historical faded colored images.Thus,there were lots of efforts trying to automate the classification operation and retrieve similar images accurately.To reach this goal,we developed a VGG19 deep convolutional neural network to extract the visual features from the images automatically.Then,the distances among the extracted features vectors are measured and a similarity score is generated using a Siamese deep neural network.The Siamese model built and trained at first from scratch but,it didn’t generated high evaluation metrices.Thus,we re-built it from VGG19 pre-trained deep learning model to generate higher evaluation metrices.Afterward,three different distance metrics combined with the Sigmoid activation function are experimented looking for the most accurate method formeasuring the similarities among the retrieved images.Reaching that the highest evaluation parameters generated using the Cosine distance metric.Moreover,the Graphics Processing Unit(GPU)utilized to run the code instead of running it on the Central Processing Unit(CPU).This step optimized the execution further since it expedited both the training and the retrieval time efficiently.After extensive experimentation,we reached satisfactory solution recording 0.98 and 0.99 F-score for the classification and for the retrieval,respectively.
基金Supported by the National Natural Science Foundation of China (Grant Nos. 60371045 & 60628101)the National High Technology Research and Development Program of China (Grant No. 2006AA01Z132)
文摘In fMRI experiments on object representation in visual cortex, we designed two types of stimuli: one is the gray face image and its line drawing, and the other is the illusion and its corresponding completed illusion. Both of them have the same global features with different minute details so that the results of fMRI experiments can be compared with each other. The first kind of visual stimuli was used in a block design fMRI experiment, and the second was used in an event-related fMRI experiment. Comparing and analyzing interesting visual cortex activity patterns and blood oxygenation level dependent (BOLD)- fMRI signal, we obtained results to show some invariance of global features of visual images. A plau- sible explanation about the invariant mechanism is related with the cooperation of synchronized re- sponse to the global features of the visual image with a feedback of shape perception from higher cortex to cortex V1, namely the integration of global features and embodiment of sparse representation and distributed population code.
基金financially supported by grants from the National Natural Science Foundation of China,No.61170136,61373101,61472270,and 61402318Natural Science Foundation(Youth Science and Technology Research Foundation)of Shanxi Province,No.2014021022-5Shanxi Provincial Key Science and Technology Projects(Agriculture),No.20130311037-4
文摘It is not clear whether the method used in functional brain-network related research can be applied to explore the feature binding mechanism of visual perception. In this study, we inves-tigated feature binding of color and shape in visual perception. Functional magnetic resonance imaging data were collected from 38 healthy volunteers at rest and while performing a visual perception task to construct brain networks active during resting and task states. Results showed that brain regions involved in visual information processing were obviously activated during the task. The components were partitioned using a greedy algorithm, indicating the visual network existed during the resting state.Z-values in the vision-related brain regions were calculated, conifrming the dynamic balance of the brain network. Connectivity between brain regions was determined, and the result showed that occipital and lingual gyri were stable brain regions in the visual system network, the parietal lobe played a very important role in the binding process of color features and shape features, and the fusiform and inferior temporal gyri were crucial for processing color and shape information. Experimental ifndings indicate that understanding visual feature binding and cognitive processes will help establish computational models of vision, improve image recognition technology, and provide a new theoretical mechanism for feature binding in visual perception.
文摘Existing glass segmentation networks have high computational complexity and large memory occupation,leading to high hardware requirements and time overheads for model inference,which is not conducive to efficiency-seeking real-time tasks such as autonomous driving.The inefficiency of the models is mainly due to employing homogeneous modules to process features of different layers.These modules require computationally intensive convolutions and weight calculation branches with numerous parameters to accommodate the differences in information across layers.We propose an efficient glass segmentation network(EGSNet)based on multi-level heterogeneous architecture and boundary awareness to balance the model performance and efficiency.EGSNet divides the feature layers from different stages into low-level understanding,semantic-level understanding,and global understanding with boundary guidance.Based on the information differences among the different layers,we further propose the multi-angle collaborative enhancement(MCE)module,which extracts the detailed information from shallow features,and the large-scale contextual feature extraction(LCFE)module to understand semantic logic through deep features.The models are trained and evaluated on the glass segmentation datasets HSO(Home-Scene-Oriented)and Trans10k-stuff,respectively,and EGSNet achieves the best efficiency and performance compared to advanced methods.In the HSO test set results,the IoU,Fβ,MAE(Mean Absolute Error),and BER(Balance Error Rate)of EGSNet are 0.804,0.847,0.084,and 0.085,and the GFLOPs(Giga Floating Point Operations Per Second)are only 27.15.Experimental results show that EGSNet significantly improves the efficiency of the glass segmentation task with better performance.
基金This work was supported by the Sichuan Science and Technology Program(2021YFQ0003).
文摘Visual question answering(VQA)has attracted more and more attention in computer vision and natural language processing.Scholars are committed to studying how to better integrate image features and text features to achieve better results in VQA tasks.Analysis of all features may cause information redundancy and heavy computational burden.Attention mechanism is a wise way to solve this problem.However,using single attention mechanism may cause incomplete concern of features.This paper improves the attention mechanism method and proposes a hybrid attention mechanism that combines the spatial attention mechanism method and the channel attention mechanism method.In the case that the attention mechanism will cause the loss of the original features,a small portion of image features were added as compensation.For the attention mechanism of text features,a selfattention mechanism was introduced,and the internal structural features of sentences were strengthened to improve the overall model.The results show that attention mechanism and feature compensation add 6.1%accuracy to multimodal low-rank bilinear pooling network.
