Schrödinger Bridge-based diffusion models have demonstrated promising performance in signal denoising. However, since ground truth signals are unavailable during the sampling process, neural networks must be employed to learn the mapping, which breaks the theoretical coupling between diffusion and sampling processes. This paper reveals a critical inconsistency between the theoretical diffusion path and the learned sampling trajectory across different frequency bands. This diffusion-sampling inconsistency directly undermines denoising effectiveness. To address this limitation, we propose the Frequency-Dependent Scheduled Schrödinger Bridge (FDSSB), which leverages power spectral density to adaptively schedule diffusion processes across frequencies. This mechanism assigns asynchronous diffusion schedules to different frequency components, correcting the diffusion schedule to better match the sampling process. As a result, FDSSB effectively mitigates the mismatch and enhances the consistency between diffusion and sampling processes. Extensive experiments demonstrate that FDSSB achieves state-of-the-art performance, with an average scale-invariant signal-to-noise ratio improvement of 7.9066 dB over competitive approaches.

Although salient object detection (SOD) methods inspired by human attention mechanisms have received increasing attention for their superior performance, they suffer from a trade-off between efficacy and efficiency, due to the high performance generally coming at the expense of large parameter sizes and high computational costs. To address this issue, we propose a Collaborative Prior-Enhanced RGB-D Salient Object Detection Network (CPENet) for Intelligent IoT Perception Devices. Specifically, we propose a dual-stream architecture utilizing MobileViT as the backbone to tackle the challenge of capturing long-range dependencies among multi-modal features with reduced model complexity. Then, a Scale Information Enhancement Module (SIEM) is proposed to effectively enhance and fuse multi-scale features. To adaptively aggregate multi-modal features, we propose a Prior-Driven Modality Aggregation (PDMA) that integrates the coarse saliency maps (Prior-Glance) of model into cross-modal fusion. Furthermore, we propose an extremely simple shared decoder to both predict the saliency maps of fused features for deep supervision and generate Prior-Glance in generation modules. Finally, considering the deficiency of boundary representation, unlike other methods which append costly edge enhancement module, a prior-driven region loss is designed that improves the capacities of boundary representation and generalization of the model under the guidance of Prior-Region. Extensive experiments demonstrate that the proposed CPENet outperforms most state-of-the-art methods on seven datasets with fewer parameters (15.7M) and lower computational complexity (21.9GFLOPs). Codes and results are available at https://github.com/blossom-lv/CPENet.

Underwater light absorption and scattering lead to severe color distortion, reduced visibility, contrast loss, and a significant degradation in image quality, thereby impeding both human visual analysis and machine vision tasks. Although considerable progress has been achieved in improving image quality, existing deep learning-based methods for underwater image enhancement (UIE) remain constrained by high computational complexity and insufficient modeling of global dependencies, which restricts their practical deployment in resource-limited underwater environments. To tackle these issues, we propose a novel hybrid framework integrating Retinex theory and state-space models (SSMs) for underwater image enhancement, named HRMamba. Different from existing Transformer-based approaches constrained by quadratic complexity, HRMamba attains computational efficiency through linear-complexity state-space operations while maintaining global dependency modeling capabilities. Moreover, to achieve comprehensive feature fusion, an Illumination Feature Fusion Module (IFFM) is proposed, which synergizes the global dependency modeling of SSMs with the local adaption capability of convolutional neural networks (CNNs). For context-sensitive noise suppression with illumination awareness, we propose an Illumination-Guided Denoising Module (IGDM) that employs directional-scanning Vision State Space Module (VSSM) blocks. Experiments demonstrate that HRMamba achieves state-of-the-art enhancement quality via an efficient architecture, significantly improving color fidelity and visibility restoration while substantially reducing computational demands. The code is available at https://github.com/YeFan-web/HRMamba/.

This article proposes a novel recognition-guided denoising method for underwater acoustic signals, termed RGDNet, which is designed to bridge the gap where existing methods, despite achieving high denoising metrics, fail to effectively enhance recognition accuracy. Specifically, we first designed a task-specific mask integration (TSMI) module that minimizes phase and frequency distortions by converting signal fusion across tasks into a unified mask-based fusion, ensuring the fusion of signal features remains effective. We further developed a context-aware signal decoupling (CASD) module to maintain the distinctiveness of sub-task signals through segmented context analysis. Following this, a recognition-guided branch (RGB) equipped with an auxiliary classifier is proposed, which guides the optimization of the denoising process based on recognition loss. Additionally, we also implemented a complex domain multimetric discriminator (CDMD) that assesses the quality of denoised signals from both metric-focused and task-specific perspectives, effectively linking denoising metrics with recognition performance. Extensive experiments show that RGDNet significantly outperforms existing methods, increasing the scale-invariant signal-to-noise ratio (SI-SNR) by an average of 6.33 dB and enhancing recognition accuracy by 11.26%.

Social recommendation improved the quality and efficiency of recommendation but increased the risk of privacy leakage, especially with the introduction of social networks. Consequently, the social recommendation considering user privacy has drawn tremendous attention from academia to industry. Nevertheless, most of the existing work regards the recommender systems as static, ignoring the diffusion of social influence over time. In this article, we propose a secure and efficient framework, temporal privacy-preserving social recommendation model (PrivTSR), to capture the changes of user preference for items and item types with time. PrivTSR first utilizes differential privacy to encrypt the data owned by the data owner. Then, inspired by the long short-term memory (LSTM), at each time step the initial user embedding and the initial item embedding are generated via DeepWalk as new ratings of users for items emerges in the user–item-type graph. The initial user-preference embedding is generated randomly at the first time step, and it is equivalent to the updated embedding of the previous time step for the later time steps. Most importantly, on the social graph, PrivTSR updates the user embedding and the user-preference embedding with graph attention convolutional network and graph attention diffused network, which aggregates (diffuses) social influence from (to) neighbors in depth and breadth. On the user–item-type graph, the user embedding and the item embedding are updated by aggregating the embedding of users and items in the six paths. Final, taking into account the users’ preference for items and item types, PrivTSR predicts the ratings of users to the items for the next time step. The extensive experiments are conducted on two real-world datasets, which demonstrated the superiority of our model over several competitive baselines.

Convolutional neural network (CNN)-based salient object detection (SOD) models have achieved promising performance in optical remote sensing images (ORSIs) in recent years. However, the restriction concerning the local sliding window operation of CNN has caused many existing CNN-based ORSI SOD models to still struggle with learning long-range relationships. To this end, a novel transformer framework is proposed for ORSI SOD, which is inspired by the powerful global dependency relationships of transformer networks. This is the first attempt to explore global and local details using transformer architecture for SOD in ORSIs. Concretely, we design an adaptive spatial tokenization transformer encoder to extract global–local features, which can accurately sparsify tokens for each input image and achieve competitive performance in ORSI SOD tasks. Then, a specific dense token aggregation decoder (DTAD) is proposed to generate saliency results, including three cascade decoders to integrate the global–local tokens and contextual dependencies. Extensive experiments indicate that the proposed model greatly surpasses 20 state-of-the-art (SOTA) SOD approaches on two standard ORSI SOD datasets under seven evaluation metrics. We also report comparison results to demonstrate the generalization capacity on the latest challenging ORSI datasets. In addition, we validate the contributions of different modules through a series of ablation analyses, especially the proposed adaptive spatial tokenization module (ASTM), which can halve the computational budget.

Since the face detection technology is widely used, improving the detection speed and accuracy in face detection tasks has become a key challenge. Therefore, this paper takes the MTCNN model as the research object and makes improvements, the purpose of which is to optimize the detection speed and accuracy simultaneously. A dynamic min size algorithm is proposed. According to the size of the input image, it dynamically controls the minimum size of the face to be recognized by the model and reduces the number of iterations of the image pyramid, increasing the number of detection frames per second by 4[Formula: see text]fps. Then, the standard convolution structure in the P-Net and R-Net models is replaced by the depth-wise separable convolution, which effectively reduces the number of parameters and computation of the model. Meanwhile, an O-Net model with a densely connected structure is also developed. Our experiments with well-known public datasets have demonstrated that the proposed network structure can improve the detection frame rate. By reusing the features of different levels of the image, the recall rate and the precision rate of the MTCNN model on the validation set are increased by 2.39% and 1.65%, respectively.