Aiming at the problem that the intermediate potential part of the traditional bistable stochastic resonance model cannot be adjusted independently, a new composite stochastic resonance(NCSR) model is proposed by combi...
详细信息
Aiming at the problem that the intermediate potential part of the traditional bistable stochastic resonance model cannot be adjusted independently, a new composite stochastic resonance(NCSR) model is proposed by combining the Woods–Saxon(WS) model and the improved piecewise bistable model. The model retains the characteristics of the independent parameters of WS model and the improved piecewise model has no output saturation, all the parameters in the new model have no coupling characteristics. Under α stable noise environment, the new model is used to detect periodic signal and aperiodic signal, the detection results indicate that the new model has higher noise utilization and better detection ***, the new model is applied to image denoising, the results showed that under the same conditions, the output peak signal-to-noise ratio(PSNR) and the correlation number of NCSR method is higher than that of other commonly used linear denoising methods and improved piecewise SR methods, the effectiveness of the new model is verified.
This paper proposes an inverse neural network approach for stochastic model calibration, focusing on the conversion of high-dimensional system sequential responses into RGB (Red, Green, and Blue) images, which signifi...
详细信息
This paper proposes an inverse neural network approach for stochastic model calibration, focusing on the conversion of high-dimensional system sequential responses into RGB (Red, Green, and Blue) images, which significantly enhances the efficiency of calibration processes. By encoding multi-nodal, multi-directional data sequence into RGB images and employing advanced neural network architectures, including the Visual Geometry Group (VGG) network for frequency response data and Long Short-Term Memory (LSTM) integrated with Residual Networks (ResNet) for sequential time-domain data, the proposed method effectively decodes complex structural responses into stochastic model parameters. This process eliminates the need for conventional iterative optimization or Bayesian sampling methods, reducing computational costs while maintaining high accuracy in parameter identification. Two case studies, the NASA Langley Uncertainty Quantification Challenge and a satellite finite element model calibration task, demonstrate the effectiveness of the approach. The novel encoding-decoding framework enables real-time model calibration for high-dimensional data, making it a promising solution for complex engineering systems with large scale, high-dimensional data and inevitable uncertainties.
Deep image Prior (DIP) has gained attention as a promising approach that bridges traditional hand-crafted priors and deep learning-based models. By utilizing a convolutional neural network (CNN) structure with "z...
详细信息
Deep image Prior (DIP) has gained attention as a promising approach that bridges traditional hand-crafted priors and deep learning-based models. By utilizing a convolutional neural network (CNN) structure with "zero sample" training, DIP effectively learns the priors of degraded images, primarily for image reconstruction. Despite its potential, there is a lack of comprehensive reviews summarizing the various DIP-based image denoising methods. This paper aims to fill this gap by providing an overview of DIP-based image denoising approaches by reviewing recent papers on the topic. We classify these methods into four groups based on their enhancement strategies: theoretical investigations, network structure, network input, and loss function. The review evaluates the strengths and weaknesses of DIP-based methods, compares state-of-the-art variants, and analyzes the impact of various improvements on denoising performance. Additionally, we identify the challenges in applying DIP to image denoising and suggest directions for future research. This review provides valuable insights into the potential of DIP in imageprocessing, especially for those new to unsupervised deep learning models.
In the era of large-scale data, the role of image compression in computer vision(CV) and computer graphics(CG) tasks is increasingly critical. Traditional methods of image compression have reached their potential limi...
详细信息
In the era of large-scale data, the role of image compression in computer vision(CV) and computer graphics(CG) tasks is increasingly critical. Traditional methods of image compression have reached their potential limits, leading to increased interest in deep learning-based techniques. However, these modern methods often compromise image quality and require extensive decoding times. This paper introduces the EICNet, which features the innovative Quick Depth-Residual Attention Module (Q-DRAM), an optimized post-processing module, and a checkerboard context model. This design aims to overcome typical shortcomings of deep learning-based compression, enhancing both training and compression efficiency as well as the quality of images at equivalent bit rates. The findings suggest that EICNet improves both the quality and efficiency of image compression. This approach marks a significant advancement in image compression technology, potentially benefiting future applications in the field.
This paper presents a convolutional neural network (CNN)-based enhancement to inter prediction in Versatile Video Coding (VVC). Our approach aims at improving the prediction signal of inter blocks with a residual CNN ...
详细信息
This paper presents a convolutional neural network (CNN)-based enhancement to inter prediction in Versatile Video Coding (VVC). Our approach aims at improving the prediction signal of inter blocks with a residual CNN that incorporates spatial and temporal reference samples. It is motivated by the theoretical consideration that neural network-based methods have a higher degree of signal adaptivity than conventional signalprocessingmethods and that spatially neighboring reference samples have the potential to improve the prediction signal by adapting it to the reconstructed signal in its immediate vicinity. We show that adding a polyphase decomposition stage to the CNN results in a significantly better trade-off between computational complexity and coding performance. Incorporating spatial reference samples in the inter prediction process is challenging: The fact that the input of the CNN for one block may depend on the output of the CNN for preceding blocks prohibits parallel processing. We solve this by introducing a novel signal plane that contains specifically constrained reference samples, enabling parallel decoding while maintaining a high compression efficiency. Overall, experimental results show average bit rate savings of 4.07% and 3.47% for the random access (RA) and low-delay B (LB) configurations of the JVET common test conditions, respectively.
The entire imagesignal Processor (ISP) of a camera relies on several processes to transform the data from the Color Filter Array (CFA) sensor, such as demosaicing, denoising, and enhancement. These processes can be e...
详细信息
The entire imagesignal Processor (ISP) of a camera relies on several processes to transform the data from the Color Filter Array (CFA) sensor, such as demosaicing, denoising, and enhancement. These processes can be executed either by some hardware or via software. In recent years, Deep Learning (DL) has emerged as one solution for some of them or even to replace the entire ISP using a single neural network for the task. In this work, we investigated several recent pieces of research in this area and provide deeper analysis and comparison among them, including results and possible points of improvement for future researchers.
To improve the imperceptibility of image steganography, an image steganography method based on a conditional invertible neural network is proposed in this paper. First, we design a conditional invertible neural networ...
详细信息
To improve the imperceptibility of image steganography, an image steganography method based on a conditional invertible neural network is proposed in this paper. First, we design a conditional invertible neural network to obtain high-quality stego images with rich high-level semantic information and clear spatial details. On the basis of the conditional directivity of the conditional invertible neural network, we can adjust the semantic information of the stego image accurately and ensure the controllability of the stego image content. We introduce a dual cross-attention module into the network structure. The integration of dual cross-attention modules enhances feature extraction and captures complex image details to improve steganographic accuracy. In addition, the introduction of the convolutional block attention module in the convolutional layer directs the model's focus to key image regions, refining stego image quality. We increase the number of convolutional blocks, which improves the ability of feature extraction and reuse. Many experiments are carried out on datasets. For the cover and stego image pairs, the PSNR value reached 43.62 dB, and for the secret and recovery image pairs, the PSNR value reached 46.48 dB. The experimental results show that the image quality and imperceptibility of this method are better than those of other state-of-the-art image steganography methods.
Deep unrolling, or unfolding, is an emerging learning-to-optimize method that unrolls a truncated iterative algorithm in the layers of a trainable neural network. However, the convergence guarantees and generalizabili...
详细信息
Deep unrolling, or unfolding, is an emerging learning-to-optimize method that unrolls a truncated iterative algorithm in the layers of a trainable neural network. However, the convergence guarantees and generalizability of the unrolled networks are still open theoretical problems. To tackle these problems, we provide deep unrolled architectures with a stochastic descent nature by imposing descending constraints during training. The descending constraints are forced layer by layer to ensure that each unrolled layer takes, on average, a descent step toward the optimum during training. We theoretically prove that the sequence constructed by the outputs of the unrolled layers is then guaranteed to converge for in-distribution problems. We then analyze the generalizability to certain out-of-distribution (OOD) shifts in the optimization problems being solved. Our analysis shows that the descending nature imposed by the proposed constraints is transferable under these distribution shifts, subject to a generalization error, thereby providing the unrolled networks with OOD robustness. We numerically assess unrolled architectures trained with the proposed constraints in two different applications, including the sparse coding using learnable iterative shrinkage and thresholding algorithm (LISTA) and image inpainting using proximal generative flow (GLOW-Prox), and demonstrate the performance and robustness advantages of the proposed method.
In many optimization problems arising from machine learning, imageprocessing, and statistics communities, the objective functions possess a special form involving huge amounts of data, which encourages the applicatio...
详细信息
In many optimization problems arising from machine learning, imageprocessing, and statistics communities, the objective functions possess a special form involving huge amounts of data, which encourages the application of stochastic algorithms. In this paper, we study such a broad class of nonconvex nonsmooth minimization problems, whose objective function is the sum of a smooth function of the entire variables and two nonsmooth functions of each variable. We propose to solve this problem with a stochastic Gauss-Seidel type inertial proximal alternating linearized minimization (denoted by SGiPALM) algorithm. We prove that under Kurdyka-Lojasiewicz (KL) property and some mild conditions, each bounded sequence generated by SGiPALM with the variance-reduced stochastic gradient estimator globally converges to a critical point after a finite number of iterations, or almost surely satisfies the finite length property. We also apply the SGiPALM algorithm to the proximal neural networks (PNN) with 4 layers for classification tasks on the MNIST dataset and compare it with other deterministic and stochastic optimization algorithms, the results illustrate the effectiveness of the proposed algorithm.
Recently, implicit neural representation (INR) has been applied to image compression. However, the rate-distortion performance of most existing INR-based image compression methods is still obviously inferior to the st...
详细信息
Recently, implicit neural representation (INR) has been applied to image compression. However, the rate-distortion performance of most existing INR-based image compression methods is still obviously inferior to the state-of-the-art image compression methods. In this letter, we propose an Enhanced Quantified Local Implicit neural Representation (EQLINR) for image compression by enhancing the utilization of local relationships of INR and narrow the quantization gap between training and encoding to further improve the performance of INR-based image compression. Our framework consists of latent representation and the corresponding implicit neural network consisting of MLP and CNN, which can transform the latent representation into the image space. To enhance local relationships utilization, we design a local enhancement module (LEM) consisted of CNN to capture the neighborhood relationships of the reconstructed image from MLP. Furthermore, to mitigate the performance loss caused by quantization of latent representation, we employ an enhanced quantization scheme (EQS) in our training process. We use uniform noise for network initialization and then use stochastic Gumbel Annealing (SGA) with dynamic temperature regulation as a proxy function for quantization during training. Extensive experimental results demonstrate that our approach significantly the compression performance of INR-based image compression, and even better than BPG.
暂无评论