Single-image desnowing aims at depressing snowflake noises while preserving a clean background. Existing methods usually mask the locations of noises and remove them in RGB color space. In this paper, we rethink this problem by investigating the impacts of color space selection. Theoretical analysis and experiments reveal that the feature of snowflake noises exhibit different distributions in different color spaces. In particular, these noises are barely seen in Hue channel, which inspires us to recover global structure and texture information of the clean background from Hue channel. More low-frequency information is also found in the Hue channel. With these observations, we propose a novel Hybrid-Color-Space-based Desnowing Network (HCSD-Net). The proposed HCSD-Net extracts low-frequency and high-frequency features in Hue channel and RGB color space, respectively. After that, it utilizes a multi-scale fusion module to enhance high-frequency details at a small feature resolution. These details are further used to supervise and supplement the background information. Extensive experiments demonstrate that our proposed HCSD-Net outperforms state-of-the-art methods on various synthetic and real-world desnowing datasets. Codes are available at https://github.com/ttz-rainbow/HCSD-Net.
论文检索
输入标题、作者或关键词，从 1,014 篇学术成果中精准定位
Satisfactory recognition performance has been achieved for simple and controllable printed molecular images. However, recognizing handwritten chemical structure images remains unresolved due to the inherent ambiguities in handwritten atoms and bonds, as well as the signifcant challenge of converting projected 2D molecular layouts into markup strings. Target to address these problems, this paper proposes an end-to-end framework for handwritten chemical structure images recognition, with novel structure-specific markup language (SSML) and random conditional guided decoder (RCGD). SSML alleviates ambiguity and complexity in Chemfig syntax by designing an innovative markup language to accurately depict molecular structures. Besides, we propose RCGD to address the issue of multiple path decoding of molecular structures, which is composed of conditional attention guidance, memory classification and path selection mechanisms. In order to fully confirm the effectiveness of the end-to-end method, a new database containing 50,000 handwritten chemical structure images (EDU-CHEMC) has been established. Experimental results demonstrate that compared to traditional SMILES sequences, our SSML can significantly reduces the semantic gap between chemical images and markup strings. It is worth noting that our method can also recognize invalid or non-existent organic molecular structures, making it highly applicable for tasks related to teaching evaluations in the fields of chemistry and biology education. The EDU-CHEMC will be released soon in https://github.com/iFLYTEK-CV/EDU-CHEMC.
In recent years, out-of-distribution (OOD) detection has emerged as a crucial research area, especially when deploying AI products in real-world scenarios. OOD detection researchers have made significant efforts to mitigate the adverse effects of abnormal activation values (abbr. activations) that refer to the outputs of the activation function acted on feature maps. Since abnormal activations would cause difficulty in separating ID and OOD data, the previous unified solution is to rectify the extremely high abnormal activations by clipping them with a pre-defined threshold or filtering them with a low-pass filter. However, it ignores the extremely low abnormal activations, and the proposed rectification strategy is always suboptimal because the used rectification function is non-convergence or high-intensity convergence, leading to under-rectification or over-rectification. In this paper, we propose an approach called Rectifying Extremely Low and High Activations (LHAct). LHAct includes a newly-designed function to rectify the extremely low and high activations at the same time. Specifically, LHAct increases the difference of means between ID and OOD activation distributions while decreasing their variances after processing the original activations. Our theoretical analyses demonstrate that LHAct significantly enhances the separability of ID and OOD data. By conducting extensive experiments, we demonstrate that LHAct surpasses previous activation-based methods significantly and generalizes well to other architectures and OOD scores. Code is available at: https://github.com/ystyuan/LHAct.git.
The recent success of deep image watermarking has demonstrated the potential of deep learning for watermarking, which has drawn increasing attention to deep video watermarking with the objective to improve its robustness and perceptual quality. Compared to images, video watermarking is much more challenging due to the rich structures of video data and the diversity of attacks in video transmission pipeline. The existing deep video watermarking schemes are far from satisfactory in dealing with temporal attacks, e.g., frame averaging, frame dropping and transcoding. To this end, a novel deep framework for Robustness Enhanced Video watermarking (REVMark) is proposed in this paper, aiming at improving the overall robustness, especially in dealing with H.264/AVC compression, while maintaining good visual quality. REVMark has an encoder/decoder structure with a pre-processing block (TAsBlock) to effectively extract the temporal-associated features on aligned frames. To ensure the end-to-end robust training, a distortion layer is integrated into the REVMark to resemble various attacks in real-world scenarios, among which, a new differentiable simulator of video compression, namely DiffH264, is developed to approximately simulate the process of H.264/AVC compression. In addition, the mask loss is incorporated to guide the encoder to embed the watermark in the human-imperceptible regions, thus improving the perceptual quality of the watermarked video. Experimental results demonstrate that the proposed scheme can outperform other SOTA methods while achieving 10X faster inference.
Deep learning-based models have demonstrated unprecedented success in image super-resolution (SR) tasks. However, more attention has been paid to lightweight SR models lately, due to the increasing demand for on-device inference. In this paper, we propose a novel Separable Modulation Network (SMN) for efficient image SR. The key parts of the SMN are the Separable Modulation Unit (SMU) and the Locality Self-enhanced Network (LSN). SMU enables global relational interactions but significantly eases the process by separating spatial modulation from channel aggregation, hence making the long-range interaction efficient. Specifically, spatial modulation extracts global contexts from spatial, and channel aggregation condenses all global context features into the channel modulator, ultimately the aggregated contexts are fused into the final features. In addition, LSN allows guiding the network to focus on more refined image attributes by encoding local contextual information. By coupling two complementary components, SMN can capture both short- and long-range contexts for accurate image reconstruction. Extensive experimental results demonstrate that our SMN achieves state-of-the-art performance among the existing efficient SR methods with less complexity.
Transformers have emerged as the superior choice for face recognition tasks, but their insufficient platform acceleration hinders their application on mobile devices. In contrast, Convolutional Neural Networks (CNNs) capitalize on hardware-compatible acceleration libraries. Consequently, it has become indispensable to preserve the distillation efficacy when transferring knowledge from a Transformer-based teacher model to a CNN-based student model, known as Cross-Architecture Knowledge Distillation (CAKD). Despite its potential, the deployment of CAKD in face recognition encounters two challenges: 1) the teacher and student share disparate spatial information for each pixel, obstructing the alignment of feature space, and 2) the teacher network is not trained in the role of a teacher, lacking proficiency in handling distillation-specific knowledge. To surmount these two constraints, 1) we first introduce a Unified Receptive Fields Mapping module (URFM) that maps pixel features of the teacher and student into local features with unified receptive fields, thereby synchronizing the pixel-wise spatial information of teacher and student. Subsequently, 2) we develop an Adaptable Prompting Teacher network (APT) that integrates prompts into the teacher, enabling it to manage distillation-specific knowledge while preserving the model's discriminative capacity. Extensive experiments on popular face benchmarks and two large-scale verification sets demonstrate the superiority of our method.
Fueled by the rapid advances in high-speed mobile networks, live video streaming has seen explosive growth in recent years and many DASH-based bitrate adaptive streaming algorithms were specifically proposed for low-latency video delivery. However, our investigations revealed that these algorithms are susceptible to network condition changes due to the use of solo universal adaptation logics, resulting the playback latency that has substantial variations across highly-fluctuating network environments and fails to meet the service quality requirement all the time. To tackle this challenge, this paper proposes Stateful Live Video Streaming (SLVS), which is a novel learning approach that learns the various network features and optimizes the adaptation logic separately for different network conditions, then dynamically tunes the logic at runtime, so that bitrate decision can better match the changing networks. Extensive evaluations show that SLVS can control playback latency down to 1s while improving Quality-of-Experience (QoE) by 17.7% to 31.8%. Moreover, it has strong robustness to maintain near-second latency over highly-fluctuating networks as well as long-period of video viewing.
With the growing demand for video applications, many advanced learned video compression methods have been developed, outperforming traditional methods in terms of objective quality metrics such as PSNR. Existing methods primarily focus on objective quality but tend to overlook perceptual quality. Directly incorporating perceptual loss into a learned video compression framework is non-trivial and raises several perceptual quality issues that need to be addressed. In this paper, we investigated these issues in learned video compression and propose a novel High Visual-Fidelity Learned Video Compression framework (HVFVC). Specifically, we design a novel confidence-based feature reconstruction method to address the issue of poor reconstruction in newly-emerged regions, which significantly improves the visual quality of the reconstruction. Furthermore, we present a periodic compensation loss to mitigate the checkerboard artifacts related to deconvolution operation and optimization. Extensive experiments have shown that the proposed HVFVC achieves excellent perceptual quality, outperforming the latest VVC standard with only 50% required bitrate.
Neural image compression has gained significant attention thanks to the remarkable success of deep neural networks. However, most existing neural image codecs focus solely on improving human vision perception. In this work, our objective is to enhance image compression methods for both human vision quality and machine vision tasks simultaneously. To achieve this, we introduce a novel approach to Partition, Transmit, Reconstruct, and Aggregate (PTRA) the latent representation of images to balance the optimizations for both aspects. By employing our method as a module in existing neural image codecs, we create a latent representation predictor that dynamically manages the bit-rate cost for machine vision tasks. To further improve the performance of auto-regressive-based coding techniques, we enhance our hyperprior network and predictor module with context modules, resulting in a reduction in bit-rate. The extensive experiments conducted on various machine vision benchmarks such as ILSVRC 2012, VOC 2007, VOC 2012, and COCO demonstrate the superiority of our newly proposed image compression framework. It outperforms existing neural image compression methods in multiple machine vision tasks including classification, segmentation, and detection, while maintaining high-quality image reconstruction for human vision.
The malicious use and widespread dissemination of deepfake pose a significant crisis of trust. Current deepfake detection models can generally recognize forgery images by training on a large dataset. However, the accuracy of detection models degrades significantly on images generated by new deepfake methods due to the difference in data distribution. To tackle this issue, we present a novel incremental learning framework that improves the generalization of deepfake detection models by continual learning from a small number of new samples. To cope with different data distributions, we propose to learn a domain-invariant representation based on supervised contrastive learning, preventing overfit to the insufficient new data. To mitigate catastrophic forgetting, we regularize our model in both feature-level and label-level based on a multi-perspective knowledge distillation approach. Finally, we propose to select both central and hard representative samples to update the replay set, which is beneficial for both domain-invariant representation learning and rehearsal-based knowledge preserving. We conduct extensive experiments on four benchmark datasets, obtaining the new state-of-the-art average forgetting rate of 7.01 and average accuracy of 85.49 on FF++, DFDC-P, DFD, and CDF2. Our code is released at \textcolorblue https://github.com/DeepFakeIL/DFIL.
TTPOINT: A Tensorized Point Cloud Network for Lightweight Action Recognition with Event Cameras
PDF ↗Event cameras have gained popularity in computer vision due to their data sparsity, high dynamic range, and low latency. As a bio-inspired sensor, event cameras generate sparse and asynchronous data, which is inherently incompatible with the traditional frame-based method. Alternatively, the point-based method can avoid additional modality transformation and naturally adapt to the sparsity of events. Still, it typically cannot reach a comparable accuracy as the frame-based method. We propose a lightweight and generalized point cloud network called TTPOINT which achieves competitive results even compared to the state-of-the-art (SOTA) frame-based method in action recognition tasks while only using 1.5 % of the computational resources. The model is adept at abstracting local and global geometry by hierarchy structure. By leveraging tensor-train compressed feature extractors, TTPOINT can be designed with minimal parameters and computational complexity. Additionally, we developed a straightforward downsampling algorithm to maintain the spatio-temporal feature. In the experiment, TTPOINT emerged as the SOTA method on three datasets while also attaining SOTA among point cloud methods on all five datasets. Moreover, by using the tensor-train decomposition method, the accuracy of the proposed TTPOINT is almost unaffected while compressing the parameter size by 55% in all five datasets.
Knowledge tracing (KT), which estimates and traces the degree of learners' mastery of concepts based on students' responses to learning resources, has become an increasingly relevant problem in intelligent education. The accuracy of predictions greatly depends on the quality of question representations. While contrastive learning has been commonly used to generate high-quality representations, the selection of positive and negative samples for knowledge tracing remains a challenge. To address this issue, we propose an adversarial bootstrapped question representation (ABQR) model, which can generate robust and high-quality question representations without requiring negative samples. Specifically, ABQR introduces the bootstrap self-supervised learning framework, which learns question representations from different views of the skill-informed question interaction graph and facilitates question representations between each view to predict one another, thereby circumventing the need for negative sample selection. Moreover, we propose a multi-objective multi-round feature adversarial graph augmentation method to obtain a higher-quality target view, while preserving the structural information of the original graph. ABQR is versatile and can be easily integrated with any base KT model as a plug-in to enhance the quality of question representation. Extensive experiments demonstrate that ABQR significantly improves the performance of the base KT model and outperforms state-of-the-art models. Ablation experiments confirm the effectiveness of each module of ABQR. The code is available at https://github.com/lilstrawberry/ABQR.
Radiance fields have gradually become a main representation of media. Although its appearance editing has been studied, how to achieve view-consistent recoloring in an efficient manner is still under explored. We present RecolorNeRF, a novel user-friendly color editing approach for the neural radiance fields. Our key idea is to decompose the scene into a set of pure-colored layers, forming a palette. By this means, color manipulation can be conducted by altering the color components of the palette directly. To support efficient palette-based editing, the color of each layer needs to be as representative as possible. In the end, the problem is formulated as an optimization problem, where the layers and their blending weights are jointly optimized with the NeRF itself. Extensive experiments show that our jointly-optimized layer decomposition can be used against multiple backbones and produce photo-realistic recolored novel-view renderings. We demonstrate that RecolorNeRF outperforms baseline methods both quantitatively and qualitatively for color editing even in complex real-world scenes.
3D dynamic point cloud (DPC) compression relies on mining its temporal context, which faces significant challenges due to DPC's sparsity and non-uniform structure. Existing methods are limited in capturing sufficient temporal dependencies. Therefore, this paper proposes a learning-based DPC compression framework via hierarchical block-matching-based inter-prediction module to compensate and compress the DPC geometry in latent space. Specifically, we propose a hierarchical motion estimation and motion compensation (Hie-ME/MC) framework for flexible inter-prediction, which dynamically selects the granularity of optical flow to encapsulate the motion information accurately. To improve the motion estimation efficiency of the proposed inter-prediction module, we further design a KNN-attention block matching (KABM) network that determines the impact of potential corresponding points based on the geometry and feature correlation. Finally, we compress the residual and the multi-scale optical flow with a fully-factorized deep entropy model. The experiment result on the MPEG-specified Owlii Dynamic Human Dynamic Point Cloud (Owlii) dataset shows that our framework outperforms the previous state-of-the-art methods and the MPEG standard V-PCC v18 in inter-frame low-delay mode.
Currently, reference-based super-resolution (RefSR) techniques leverage high-resolution (HR) reference images to provide useful content and texture information for low-resolution (LR) images during the super-resolution (SR) process. Nevertheless, it is time-consuming, laborious, and even impossible in some cases to find high-quality reference images. To tackle this problem, we propose a brand-new self-reference image super-resolution approach using a pre-trained diffusion large model and a window adjustable transformer, termed DWTrans. Our proposed method does not require explicitly inputting manually acquired reference images during training and inference. Specifically, we feed the degraded LR images into a pre-trained stable diffusion large model to automatically generate corresponding high-quality self-reference (SRef) images that provide valuable high-frequency details for the LR images in the process of SR. To extract valuable high-frequency information in SRef images, we design a window adjustable transformer with both non-adjustable window layer (NWL) and adjustable window layer (AWL). The NWL learns local features from LR images using a dense window, while the AWL acquires global features from the SRef images using a random sparse window. Furthermore, to fully utilize the high-frequency features in the SRef image, we introduce the adaptive deformable fusion module to adaptively fuse the features of the LR and SRef images. Experimental results validate that our proposed DWTrans outperforms state-of-the-art methods on various benchmark datasets both quantitatively and visually.
SEAM: Searching Transferable Mixed-Precision Quantization Policy through Large Margin Regularization
PDF ↗Mixed-precision quantization (MPQ) suffers from the time-consuming process of searching the optimal bit-width allocation (i.e., the policy) for each layer, especially when using large-scale datasets such as ISLVRC-2012. This limits the practicality of MPQ in real-world deployment scenarios. To address this issue, this paper proposes a novel method for efficiently searching for effective MPQ policies using a small proxy dataset instead of the large-scale dataset used for training the model. Deviating from the established norm of employing a consistent dataset for both model training and MPQ policy search stages, our approach, therefore, yields a substantial enhancement in the efficiency of MPQ exploration. Nonetheless, using discrepant datasets poses challenges in searching for a transferable MPQ policy. Driven by the observation that quantization noise of sub-optimal policy exerts a detrimental influence on the discriminability of feature representations---manifesting as diminished class margins and ambiguous decision boundaries---our method aims to identify policies that uphold the discriminative nature of feature representations, i.e., intra-class compactness and inter-class separation. This general and dataset-independent property makes us search for the MPQ policy over a rather small-scale proxy dataset and then the policy can be directly used to quantize the model trained on a large-scale dataset. Our method offers several advantages, including high proxy data utilization, no excessive hyper-parameter tuning, and high searching efficiency. We search high-quality MPQ policies with the proxy dataset that has only 4% of the data scale compared to the large-scale target dataset, achieving the same accuracy as searching directly on the latter, improving MPQ searching efficiency by up to 300×.
Towards Real-Time Neural Video Codec for Cross-Platform Application Using Calibration Information
PDF ↗The state-of-the-art neural video codecs have outperformed the most sophisticated traditional codecs in terms of rate-distortion (RD) performance in certain cases. However, utilizing them for practical applications is still challenging for two major reasons. 1) Cross-platform computational errors resulting from floating point operations can lead to inaccurate decoding of the bitstream. 2) The high computational complexity of the encoding and decoding process poses a challenge in achieving real-time performance. In this paper, we propose a real-time cross-platform neural video codec, which is capable of efficiently decoding (25FPS) of 720P video bitstream from other encoding platforms on a consumer-grade GPU (e.g., NVIDIA RTX 2080). First, to solve the problem of inconsistency of codec caused by the uncertainty of floating point calculations across platforms, we design a calibration transmitting system to guarantee the consistent quantization of entropy parameters between the encoding and decoding stages. The parameters that may have transboundary quantization between encoding and decoding are identified in the encoding stage, and their coordinates will be delivered by auxiliary transmitted bitstream. By doing so, these inconsistent parameters can be processed properly in the decoding stage. Furthermore, to reduce the bitrate of the auxiliary bitstream, we rectify the distribution of entropy parameters using a piecewise Gaussian constraint. Second, to match the computational limitations on the decoding side for real-time video codec, we design a lightweight model. A series of efficiency techniques, such as model pruning, motion downsampling, and arithmetic coding skipping, enable our model to achieve 25 FPS decoding speed on NVIDIA RTX 2080 GPU. Experimental results demonstrate that our model can achieve real-time decoding of 720P videos while encoding on another platform. Furthermore, the real-time model brings up to a maximum of 24.2% BD-rate improvement from the perspective of PSNR with the anchor H.265 (medium).
Recent years have witnessed a few attempts of vision transformers for single image super-resolution (SISR). Since the high resolution of intermediate features in SISR models increases memory and computational requirements, efficient SISR transformers are more favored. Based on some popular transformer backbone, many methods have explored reasonable schemes to reduce the computational complexity of the self-attention module while achieving impressive performance. However, these methods only focus on the performance on the training platform (e.g., Pytorch/Tensorflow) without further optimization for the deployment platform (e.g., TensorRT). Therefore, they inevitably contain some redundant operators, posing challenges for subsequent deployment in real-world applications. In this paper, we propose a deployment-friendly transformer unit, namely UFONE (i.e., UnFolding ONce is Enough), to alleviate these problems. In each UFONE, we introduce an Inner-patch Transformer Layer (ITL) to efficiently reconstruct the local structural information from patches and a Spatial-Aware Layer (SAL) to exploit the long-range dependencies between patches. Based on UFONE, we propose a Deployment-friendly Inner-patch Transformer Network (DITN) for the SISR task, which can achieve favorable performance with low latency and memory usage on both training and deployment platforms. Furthermore, to further boost the deployment efficiency of the proposed DITN on TensorRT, we also provide an efficient substitution for layer normalization and propose a fusion optimization strategy for specific operators. Extensive experiments show that our models can achieve competitive results in terms of qualitative and quantitative performance with high deployment efficiency.
Image steganography is the art of concealing secret information in images in a way that is imperceptible to unauthorized parties. Recent advances show that is possible to use a fixed neural network (FNN) for secret embedding and extraction. Such fixed neural network steganography (FNNS) achieves high steganographic performance without training the networks, which could be more useful in real-world applications. However, the existing FNNS schemes are vulnerable in the sense that anyone can extract the secret from the stego-image. To deal with this issue, we propose a key-based FNNS scheme to improve the security of the FNNS, where we generate key-controlled perturbations from the FNN for data embedding. As such, only the receiver who possesses the key is able to correctly extract the secret from the stego-image using the FNN. In order to improve the visual quality and undetectability of the stego-image, we further propose an adaptive perturbation optimization strategy by taking the perturbation cost into account. Experimental results show that our proposed scheme is capable of preventing unauthorized secret extraction from the stego-images. Furthermore, our scheme is able to generate stego-images with higher visual quality than the state-of-the-art FNNS scheme, especially when the FNN is a neural network for ordinary learning tasks.
Inspired by certain optimization solvers, the deep unfolding network (DUN) usually inherits a multi-phase structure for image compressive sensing (CS). However, in existing DUNs, the message transmission within and between phases still faces two issues: 1) the roughness of transmitted information, e.g., the low-dimensional representations. 2) the inefficiency of transmitted policy, e.g., simply concatenating deep features. In this paper, by unfolding the Proximal Gradient Descent (PGD) algorithm, a novel gradient guided DUN (G2 -DUN) for image CS is proposed, in which a gradient map is delicately introduced within each phase for providing richer informational guidance at both intra-phase and inter-phase levels. Specifically, corresponding to the gradient descent (GD) of PGD, a gradient guided GD module is designed, in which the gradient map can adaptively guide step size allocation for different textures of input image, realizing a content-aware gradient updating. On the other hand, corresponding to the proximal mapping (PM) of PGD, a gradient guided PM module is developed, in which the gradient map can dynamically guide the exploring of deep textural priors in multi-scale space, achieving the dynamic perception of the proposed deep model. By introducing the gradient map, the proposed message transmission system not only facilitates the informational communication between different functional modules within each phase, but also strengthens the inferential cooperation among cascaded phases. Extensive experiments manifest that the proposed G2 -DUN outperforms existing state-of-the-art CS methods.