Continual semantic segmentation (CSS) has risen as a popular field, which aims to acquire new skills constantly without forgetting past knowledge catastrophically. In CSS, we identify that there is a severe imbalance ...
详细信息
ISBN:
(数字)9798350390155
ISBN:
(纸本)9798350390162
Continual semantic segmentation (CSS) has risen as a popular field, which aims to acquire new skills constantly without forgetting past knowledge catastrophically. In CSS, we identify that there is a severe imbalance between new classes and old classes, leading to the classifier weight toward new classes. In this paper, we deal with the continual semantic segmentation problem from the class imbalance perspective via mask-based class rebalancing, avoiding the model suffering from catastrophic forgetting. More specifically, the mask-based class rebalancing depends on a mask to combine resampling with reweighting ingenuously, which mitigates the classifier bias toward new classes. Besides, we also propose a frequency knowledge distillation, leveraging multiple frequency components information to maintain the feature representation space for old classes. We demonstrate the effectiveness of our approach with an extensive evaluation of the Pascal-VOC 2012 and ADE20K datasets, significantly outperforming the state-of-the-art method.
Light field, as a new data representation format in multimedia, has the ability to capture both intensity and direction of light rays. However, the additional angular information also brings a large volume of data. Cl...
详细信息
ISBN:
(纸本)9781665492584
Light field, as a new data representation format in multimedia, has the ability to capture both intensity and direction of light rays. However, the additional angular information also brings a large volume of data. Classical coding methods are not effective to describe the relationship between different views, leading to redundancy left. To address this problem, we propose a novel light field compression scheme based on implicit neural representation to reduce redundancies between views. We store the information of a light field image implicitly in an neural network and adopt model compression methods to further compress the implicit representation. Extensive experiments have demonstrated the effectiveness of our proposed method, which achieves comparable rate-distortion performance as well as superior perceptual quality over traditional methods.
Over the past few years, the advancement of Multimodal Large Language Models (MLLMs) has captured the wide interest of researchers, leading to numerous innovations to enhance MLLMs’ comprehension. In this paper, we p...
详细信息
With the success of generative adversarial networks (GANs) on various real-world applications, the controllability and security of GANs have raised more and more concerns from the community. Specifically, understandin...
详细信息
Synthetic aperture radar (SAR) tomography (TomoSAR) has garnered significant attention due to its capability for three-dimensional reconstruction. Compressed sensing (CS) methods are widely employed to address the Tom...
ISBN:
(数字)9781837240982
Synthetic aperture radar (SAR) tomography (TomoSAR) has garnered significant attention due to its capability for three-dimensional reconstruction. Compressed sensing (CS) methods are widely employed to address the TomoSAR inversion challenge. Nevertheless, practical applications reveal phase errors among different channels, resulting in defocusing and blurring when relying solely on CS for 3D reconstruction. Current state-of-the-art autofocus techniques suffer from prohibitive computational complexity, limiting their applicability to large-scale 3D imaging. In pursuit of efficient TomoSAR 3-D autofocusing, we proposed ASAMP-Net, an innovative deep unfolding network. Operating within a two-step framework, each layer comprises two stages: phase error estimation and iterative scattering coefficient reconstruction using the sparse adaptive matching pursuit (SAMP) algorithm. Additionally, phase error estimation is obtained through mathematical derivation, while challenges associated with fixed sparsity and limited efficiency in conventional methods are mitigated through deep learning techniques. Simulation experiments and real data validation affirm the effectiveness and superiority of the proposed method.
The synthetic aperture radar (SAR) can be affected by various types of jamming during operation. Among them, the deceptive jamming generated by digital radio frequency memory (DRFM) jammers poses a serious threat to S...
详细信息
The synthetic aperture radar (SAR) can be affected by various types of jamming during operation. Among them, the deceptive jamming generated by digital radio frequency memory (DRFM) jammers poses a serious threat to SAR imaging by creating highly realistic false targets. Moreover, with advancements in deceptive jamming technology, the generation speed of deceptive jamming has increased, rendering existing methods less effective. To address this issue, an anti-deceptive jamming method based on pulse repetition interval (PRI) variation design and multi-channel principle is proposed to mitigate the effects of deceptive jamming. First, a PRI variation strategy that will not cause the loss of echo signals in the imaging area is designed. By utilizing this strategy for imaging, deceptive jamming signals are dispersed across different ranges, resulting in preliminary suppression of the jamming. Subsequently, after azimuth non-uniform sampling reconstruction and range processing, most of the jamming signals are suppressed due to the azimuth timing differences between SAR and jamming signals. However, when the jammer uses specific retransmission intervals, such as the average PRI of the PRI sequence, the jamming signals may be concentrated at certain ranges, retaining some coherence and posing a threat to SAR imaging. To overcome this challenge, a residual jamming detection and suppression algorithm based on multi-channel principle is proposed, which can detect and filter out the channels affected by jamming. Finally, an azimuth sparse reconstruction is introduced for azimuth processing. Since the anti-jamming principle of this method relies on the differences in azimuth timing between SAR and jamming, it can suppress deceptive jamming even when the generation speed of deceptive jamming is rapid, which some other anti-deceptive jamming methods cannot achieve. Simulations of SAR imaging under deceptive jamming conditions are conducted for point target scene and complex target
On-orbit processing is becoming more prevalent due to its ability to efficiently exploit satellite resources. On-orbit geometric rectification improves positioning accuracy for follow-up tasks such as object detection...
On-orbit processing is becoming more prevalent due to its ability to efficiently exploit satellite resources. On-orbit geometric rectification improves positioning accuracy for follow-up tasks such as object detection or geometric calibration, while avoiding heavy burden on downlinking bandwidth and time delay. However, existing rectification methods faces some challenges. The hardware resources onboard satellites are restricted, and geographic positioning is often inaccurate. In this article, we propose a novel method designed for on-orbit rectification. The proposed method introduces a two-step registration framework to overcome large initial offsets and also a feature-compressing strategy to reduce the storage space of reference patches. Quantitative and practical experiments demonstrate that the proposed method performs well in terms of storage space, time efficiency as well as registration accuracy.
Deep neural networks (DNNs) have shown great potential in non-reference image quality assessment (NR-IQA). However, the annotation of NR-IQA is labor-intensive and time-consuming, which severely limits their applicati...
详细信息
Existing learning-based methods for blind image quality assessment (BIQA) are heavily dependent on large amounts of annotated training data, and usually suffer from a severe performance degradation when encountering t...
详细信息
With the increase in the number of remote sensing satellites and imaging modes, the amount of data for acquiring remote sensing images has greatly increased. Effectively and stably performing geometric positioning on ...
详细信息
ISBN:
(数字)9798350385991
ISBN:
(纸本)9798350386004
With the increase in the number of remote sensing satellites and imaging modes, the amount of data for acquiring remote sensing images has greatly increased. Effectively and stably performing geometric positioning on remote sensing images is the foundation of remote sensing applications. This paper proposes a remote sensing image matching and positioning method based on a multi-feature control point database in MySQL. Firstly, a feature control point database in MySQL is constructed based on multiple feature methods. Subsequently, the target image is matched from coarse to fine using region features and point features in the feature control point database. Experimental results show that, on three target remote sensing images, the coarse-to-fine matching method based on MySQL multi-feature database can achieve good geometric positioning effects, with a positioning accuracy of around 0.5 pixels.
暂无评论