This paper presents a review of the second LoViF Challenge on Real-World All-in-One Image Restoration. The challenge aims to advance unified image restoration under diverse real-world degradation conditions, including blur, low-light, haze, rain, and snow. It provides a common benchmark for evaluating the restoration accuracy, robustness, and generalization capability of models across multiple degradation categories within a unified framework. The competition attracted 158 registered participants, and 20 teams were included in the final ranking after their submitted results were successfully reproduced and verified. This report provides a comprehensive analysis of the submitted solutions and corresponding results, highlighting recent advances in real-world all-in-one image restoration. The summarized methods and empirical findings reveal effective design strategies and establish an updated benchmark for future research in real-world low-level vision.

The smooth and safe operation of underground conveyor belts is a fundamental prerequisite for efficient and secure coal-mine production. Manual inspection is inefficient and prone to missed detections; therefore, foreign objects upon entering the conveyor belt must be identified accurately and promptly. Due to the dim underground environment, dust interference, and low image contrast, conventional visual processing methods cannot satisfy the requirements of this application scenario. To address these problems, this study develops a lightweight YOLOv8-based detection framework in which Contrast Limited Adaptive Histogram Equalization (CLAHE) is first used as an image-preprocessing strategy to enhance local contrast and suppress noise. The network is then optimized by replacing selected standard convolutions with Alterable Kernel Convolution (AKConv), which was selected because its adaptive sampling offsets and linear parameter scaling are better suited to elongated and irregular mining debris than fixed square kernels or heavier deformable-convolution variants. Compared with the baseline YOLOv8, the proposed model increases Precision from 86.3

Effective long-term predictions are increasingly sought in urban data mining systems. Existing methods typically handle spatial and temporal dependencies in isolation, failing to integrate long-term trends with local correlations. To address this issue, we propose the Trend-Local Convolutional Network (TLCN), which consists of three components: Cross-Period Sparse Downsampling (CPSD), Trend Convolution (TC), and Local Convolution (LC). TLCN model can not only learn long-term dependencies through TC, but also learn local spatial dependencies through Local Convolution. In addition, the intermediate step between the previous and next time steps is removed in TLCN model, which leads to a decrease in the complexity. Extensive experiments on four real-world datasets validate the superior performance of TLCN. The experimental results demonstrate that the proposed TLCN model outperforms existing state-of-the-art models by 0.56 and 0.37 on PEMS08, and by 0.46 and 0.43 on METR-LA across the RMSE and MAPE metrics.

Recently, GRPO-based reinforcement learning has shown remarkable progress in optimizing flow-matching models, effectively improving their alignment with task-specific rewards. Within these frameworks, the policy update relies on importance-ratio clipping to constrain overconfident positive and negative gradients. However, in practice, we observe a systematic shift in the importance-ratio distribution—its mean falls below 1 and its variance differs substantially across timesteps. This left-shifted and inconsistent distribution prevents positive-advantage samples from entering the clipped region, causing the mechanism to fail in constraining overconfident positive updates. As a result, the policy model inevitably enters an implicit over-optimization stage —while the proxy reward continues to increase, essential metrics such as image quality and text–prompt alignment deteriorate sharply, ultimately making the learned policy impractical for real-world use. To address this issue, we introduce GRPO-Guard , a simple yet effective enhancement to existing GRPO frameworks. Our method incorporates ratio normalization, which restores a balanced and step-consistent importance ratio, ensuring that PPO clipping properly constrains harmful updates across denoising timesteps. In addition, a gradient reweighting strategy equalizes policy gradients over noise conditions, preventing excessive updates from particular timestep regions. Together, these designs act as a regulated clipping mechanism, stabilizing optimization and substantially mitigating implicit over-optimization without relying on heavy KL regularization. Extensive experiments on multiple diffusion backbones (e.g., SD3.5M, Flux.1-dev) and diverse proxy tasks demonstrate that GRPO-Guard significantly reduces over-optimization while maintaining or even improving generation quality. We provide detailed demonstrations of the over-optimization process and corresponding visualizations in Supplementary Materials. 5 .

The rapid advancement of AI-generated content (AIGC) makes robust watermarking essential for content authenticity and preventing misuse. Local watermarking, which embeds watermarks in specific image regions, is noted for strong cropping resistance and low visual distortion. However, such methods rely on keypoints for locating embedding regions, which are prone to drift or disappearance under attacks, leading to unreliable extraction. To enhance local watermarking robustness, we propose a method that simultaneously performs watermark embedding and keypoint enhancement through a single adaptive pixel modification, preventing interference. The superiority of our method is reflected in three aspects: (1) A joint optimization function is constructed to determine the optimal pixel modification that embeds the watermark while enhancing keypoint strength, improving resistance to attacks. (2) A mapping between the spatial and frequency domains allows direct modification of frequency coefficients in the spatial domain to embed the watermark, enhancing efficiency. (3) The overall modification is constrained by the Peak Signal-to-Noise Ratio (PSNR) and allocated pixel-wise based on the Just Noticeable Difference (JND) to ensure imperceptibility. Experimental results show our method achieves superior robustness over existing traditional and deep watermarking methods, while maintaining high visual quality with a PSNR above 40 dB under various typical attacks.

With the rapid advancement of generative models, high-fidelity AI-generated images have become increasingly indistinguishable from real images, posing significant challenges to traditional detection methods that rely on explicit artifacts or uniform feature learning. We hypothesize that detection ambiguity originates from pattern coexistence: synthetic images simultaneously embed (a) authentic patterns inherited from real-image distributions and (b) synthetic patterns induced by generative architectures, whereas real images maintain consistent patterns. We validate this hypothesis through SHAP-based quantitative analysis, demonstrating that synthetic images inherently exhibit a dual distribution—simultaneously containing authentic patterns and synthetic traces—while real images show a unimodal distribution. Building on this insight, this paper proposes a Dual-Branch Asymmetric Discrepancy Learning (DADL) framework. The DADL leverages multi-scale feature extraction and Asymmetric Feature Discrepancy Loss to capture and amplify such pattern differences across multiple scales. Extensive experiments on three benchmarks (AIGCDetectBenchmark, GenImage, and Chameleon) show that DADL achieves state-of-the-art performance, with particular strengths in detecting high-fidelity synthetic images from diffusion models (e.g., Midjourney, SDv1.4, SDv1.5) and enhancing generalization across diverse generative paradigms. This study not only offers an effective approach for AIGI detection but also sheds light on the intrinsic properties of synthetic images, providing a new perspective for advancing AIGI forensics.

Deep learning-based image inpainting methods can synthesise plausible results using information from undamaged regions of the input image. However, most existing methods fail to generate high-quality structures and textures due to insufficient exploration of high-level structure-aware information. To address this issue, a cascaded 'channel-first and spatial-second' attention module that prioritises structural enhancement before texture matching is proposed, ensuring structure-aware coherence in restored regions. And then, a localised hybrid attention mechanism that operates only on high-level features is designed, achieving significant efficiency gains. Finally, this mechanism is integrated into a lightweight single-stream GAN, reducing parameters by 82% and FLOPs by 71% compared to representative two-stage models while maintaining competitive inpainting quality. Experiments on multiple datasets, including faces and natural images, demonstrate that higher-quality results with more details are generated by the proposed SGHA-GAN compared to state-of-the-art methods.

Transformer-based models have demonstrated state-of-theart results in the field of image super-resolution. However, we observe that such methods sometimes suffer from overly smooth structural reconstruction and blurred details, indicating that the potential of Transformers has not yet been fully exploited in existing networks. To leverage more prior information, this paper proposes a novel Global Cross Attention Transformer (GCAT) algorithm. This algorithm introduces external prior information by incorporating a cross-attention mechanism alongside the original self-attention mechanism. Furthermore, to better establish the model, we apply cross-attention across all Transformer modules to enhance the model capability for complex mapping. Extensive experiments demonstrate the efficacy of the proposed architecture, with the overall approach exceeding the performance of current state-of-the-art methods.

Current cultural relic image inpainting methods mainly utilize single encoder-decoder architectures. However, single encoder-decoder methods struggle with introducing prior conditions, especially for historical relics datasets with unique damage patterns. A Multi-column Condition Decoding Transformer for Cultural Relic Image Inpainting (MCDT) is proposed to address above issue. The proposed MCDT model employs multi-column decoders integrated into the Transformer through cross-attention mechanism. Specifically, multi-column decoders consist of three branches: (1) a self-attention branch that decodes the encoded latent features, (2) a ground-truth cross-attention branch that enforces constraints from ground truth data, and (3) an edge cross-attention branch that incorporates edge constraints. The multi-column decoding architecture enables the simultaneous integration of multiple external conditions to constitute a multi-prior constrained image inpainting model. Comparative experiments conducted on cultural relic dataset show that the proposed MCDT method generates higher-quality inpainting results compared to state-of-the-art methods.

Image super-resolution (SR) plays a vital role in vision tasks, in which Transformer-based methods outperform conventional convolutional neural networks. Existing work usually uses residual linking to improve the performance, but this type of linking provides limited information transfer within the block. Also, existing work usually restricts the self-attention computation to a single window to improve feature extraction. This means transformer-based networks can only use feature information within a limited spatial range. To handle the challenge, this paper proposes a novel Hybrid Attention-Dense Connected Transformer Network (HADT) to utilize the potential feature information better. HADT is constructed by stacking an attentional transformer block (ATB), which contains an Effective Dense Transformer Block (EDTB) and a Hybrid Attention Block (HAB). EDTB combines dense connectivity and swin-transformer to enhance feature transfer and improve model representation, and meanwhile, HAB is used for cross-window information interaction and joint modeling of features for better visualization. Based on the experiments, our method is effective on SR tasks with magnification factors of 2, 3, and 4. For example, using the Urban100 dataset in an experiment with an amplification factor of 4 our method has a PSNR value that is 0.15 dB higher than the previous method and reconstructs a more detailed texture.

This article discusses our team's success in securing first place in ICPR 2024 Multi-Modal Visual Pattern Recognition ChallengeTrack 2 by applying data augmentation, focal loss, and model ensemble techniques. Data enhancement improves training data diversity, focal loss addresses class imbalance by focusing on challenging examples, and model ensembling combines predictions from multiple models for better performance. Together, these strategies lead to significant improvements in model accuracy and robustness, contributing to our high-ranking result. Code at https://github.com/chaoyuhao/ICPR24 competition.git.

Extraction of agricultural field parcels is of great importance for agricultural condition monitoring, farm management, and food security. Several methods have been developed to map the distribution of agricultural field parcels, among which deep learning-based supervised learning is increasingly employed. Nevertheless, advanced deep learning models face two major limitations: limited ability to generalize across different spatial,temporal and sensor contexts with varying scene and object characteristics, and high requirement for annotated datasets to support training and validation. To address this challenge, we introduce a novel unsupervised domain adaptation (UDA) framework (UDA-Field Teacher, UDA-FT) for agricultural field parcel instance segmentation, which is designed to transfer knowledge from labeled source domains to unlabeled target domains. UDA-FT is based on the Mask R-CNN framework and incorporates a target-oriented teacher model and a cross-domain student model. This cross-domain student model embeds an image adaptation module and an instance adaptation module, employing adversarial learning strategies to mitigate cross-domain distribution differences. Additionally, we propose a consistency mutual learning module based on soft pseudo-label technology, overcoming the limitations of traditional hard pseudo-labeling in confidence threshold selection and improving model robustness in the target domain. Furthermore, to address the difficulty in generating independent instance labels for densely packed agricultural field parcels and capturing spatial contextual relationships during soft pseudo-label generation, we propose two data augmentation methods, namely CutMatch (CM) and LeakyMask (LM). We adopted the proposed framework on cross-scene and cross-sensor datasets to evaluate its effectiveness and robustness under different scenes. Quantification and visualization results demonstrate our UDA-FT outperforms existing domain adaptation methods for cross-scene and cross-sensor agricultural field parcels across all metrics. Ablation studies highlight the substantial impact of strong data augmentation on model performance, emphasizing the importance of learning from out-of-distribution data. As an innovative application of unsupervised domain adaptation in agricultural field parcel instance segmentation, this research provides a novel method for domain shift in agricultural remote sensing imagery, enabling more accurate field instance segmentation with significant implications for global agriculture.

Existing deep image watermarking methods follow a fixed embedding-distortion-extraction pipeline, where the embedder and extractor are weakly coupled through a final loss and optimized in isolation. This design lacks explicit collaboration, leaving no structured mechanism for the embedder to incorporate decoding-aware cues or for the extractor to guide embedding during training. To address this architectural limitation, we rethink deep image watermarking by reformulating embedding and extraction as explicitly collaborative components. To realize this reformulation, we introduce a Collaborative Interaction Mechanism (CIM) that establishes direct, bidirectional communication between the embedder and extractor, enabling a mutual-teacher training paradigm and coordinated optimization. Built upon this explicitly collaborative architecture, we further propose an Adaptive Feature Modulation Module (AFMM) to support effective interaction. AFMM enables content-aware feature regulation by decoupling modulation structure and strength, guiding watermark embedding toward stable image features while suppressing host interference during extraction. Under CIM, the AFMMs on both sides form a closed-loop collaboration that aligns embedding behavior with extraction objectives. This architecture-level redesign changes how robustness is learned in watermarking systems. Rather than relying on exhaustive distortion simulation, robustness emerges from coordinated representation learning between embedding and extraction. Experiments on real-world and AI-generated datasets demonstrate that the proposed method consistently outperforms state-of-the-art approaches in watermark extraction accuracy while maintaining high perceptual quality, showing strong robustness and generalization.

Video object removal and inpainting are critical tasks in the fields of computer vision and multimedia processing, aimed at restoring missing or corrupted regions in video sequences. Traditional methods predominantly rely on flow-based propagation and spatio-temporal Transformers, but these approaches face limitations in effectively leveraging long-term temporal features and ensuring temporal consistency in the completion results, particularly when dealing with large masks. Consequently, performance on extensive masked areas remains suboptimal. To address these challenges, this paper introduces a novel video inpainting approach leveraging the Diffusion Transformer (DiT). DiT synergistically combines the advantages of diffusion models and transformer architectures to maintain long-term temporal consistency while ensuring high-quality inpainting results. We propose a Circular Position-Shift strategy to further enhance long-term temporal consistency during the inference stage. Additionally, the proposed method interactively removes specified objects, and generates corresponding prompts. In terms of processing speed, it takes only 65 seconds (testing on one NVIDIA H800 GPU) to complete a video with a resolution of 2160 × 2100 with 97 frames without any acceleration method. Experimental results indicate that the proposed method demonstrates superior performance in content fidelity, texture restoration, and temporal consistency. Project page:https://jieliu95.github.io/EraserDiT_demo/

Tensor decomposition can effectively reduce redundant parameters and FLOPs in deep neural network (DNN). DNNs with low parameters and computational complexity are more convenient to deploy on edge devices. As a combination of CANDECOMP/PARAFAC decomposition and Tucker decomposition, block term decomposition (BTD) demonstrates stronger generalization performance but presents challenges in configuring the number of Tucker tensors and the dimensions of each core tensor, making the tuning of decomposition rank more complex. To solve the problem of complex rank configuration in BTD, a novel block term decomposition of deep neural networks by variational Bayesian (BTD-VB) is proposed. BTD-VB mainly consists of three parts: selecting the rank of the network layer parameter tensor, compressing network parameters through BTD, and fine-tuning compressed model. The BTD-VB algorithm automatically estimates the rank of DNN parameters in input and output channel dimensions through the global analytic solution of empirical variational Bayesian matrix factorization. Extensive experimental results show that compared with state-of-the-art methods, BTD-VB achieves a higher compression rate while guaranteeing compression performance.

Video has become a core medium for information dissemination and entertainment, but the surging demand has led to frequent infringement, posing severe challenges to copyright protection. Existing video watermarking methods face limitations such as quality degradation, neglect of encoding/decoding impacts, and lack of audio copyright protection. To address these, we propose an FFmpeg-based audio-visual dual watermarking framework deeply integrated into the video encoding/decoding pipeline. Our framework employs the joint selection method of interval k frames and key frames based on DINOv2 for image embedding, which leverages DINOv2's self-supervised features to jointly optimize frame selection, while adopting averageamplitude threshold screening for audio segments to ensure watermark imperceptibility. In addition, we propose a twostage voting verification that optimizes extraction accuracy via confidence voting. Our approach achieves dual-watermark crossvalidation, significantly enhancing robustness against complex attacks. Experimental results demonstrate that the method achieves a 0 Bit Error Rate (BER) under various attack scenarios, providing an innovative solution for multimodal digital copyright protection with notable theoretical and practical value.

Tensor ring (TR) decomposition is an effective method to achieve deep neural network (DNN) compression. However, there are two problems with TR decomposition: setting TR rank to equal in TR decomposition and selecting rank through an iterative process is time-consuming. To address the two problems, A TR network compression method by Bayesian optimization (TR-BO) is proposed. TR-BO involves selecting rank via Bayesian optimization, compressing the neural network layer via TR decomposition using rank obtained in the previous step, and, finally, further fine-tuning the compressed model to overcome some of the performance loss due to compression. Experimental results show that TR-BO achieves the best results in terms of Top-1 accuracy, parameter, and training time. For example, on the CIFAR-10 dataset Resnet20 network, TR-BO-1 achieves 87.67% accuracy with a compression ratio of 13.66 and a running time of only 2.4 hours. Furthermore, TR-BO has achieved state-of-the-art performance on the CIFAR-10/100 benchmark tests.

At present, two-stage networks are widely used in image restoration methods, but existing two-stage network often generates inpainting results with distorted structures and blurry textures, especially when reconstructed object is more complex. The main reason is insufficient structure prior and inaccurate, which leads to generating wrong results in texture generation stage. In order to solve this problem, a novel Image Inpainting based on Edge and Smooth Structures Prediction is proposed. The edge structure and smooth structure are completed in structure reconstruction stage, and reconstructed edge structure and smooth structure are simultaneously used as a prior to guide texture generation stage fills in damaged area. The proposed method is evaluated on publicly available datasets Paris StreetView, CelebA-HQ and Places2, and many experiments show that proposed method obtains excellent results under subjective and objective indexes compared with mainstream approaches.