Jing Wang

To address the challenge of satisfying strict Service-Level Agreement (SLA) requirements for concurrent smart video surveillance tasks in heterogeneous edge computing environments, an SLA-aware adaptive scheduling algorithm for heterogeneous computing collaboration is proposed. First, a mixed-task flow model is constructed, and a finite-state Markov chain is utilized to dynamically model the time-varying wireless channel. Second, a Dueling Double Deep Q-Network (Dueling DDQN) scheduling algorithm based on SLA awareness and channel adaptation is proposed, with a designed SLA action-masking mechanism. This mechanism advances hard delay constraints to the decision-generation stage, dynamically prunes the action space based on real-time channel conditions and node loads, and filters out actions predicted to violate the SLA before execution. Experimental results show that the proposed algorithm coordinates heterogeneous computing resources between the cloud center and the edge and exhibits earlier empirical reward stabilization and lower task-violation rates than the compared learning-based baselines under the tested workload conditions.

Joint semantic segmentation and monocular depth estimation are essential for environmental perception of Autonomous Underwater Vehicles (AUVs), yet achieving high-precision real-time multi-task perception remains challenging due to severe underwater optical degradation and the limited computational resources of AUV platforms. This paper proposes an efficient MobileMamba-Based Multi-Task Learning Network (MM-MTL) for underwater scene perception. The network employs MobileMamba, a lightweight State Space Model (SSM) architecture, as the backbone to balance prediction accuracy and computational efficiency. An efficient multi-task collaborative decoder is further designed, comprising a Local Detail Enhancement (LDE) module that sharpens edge features in degraded underwater images via multi-scale parallel pooling, and a Dynamic Attention Fusion Strategy (DAFS) that adaptively fuses semantic and depth features across channel and spatial dimensions, exploiting cross-task synergies while suppressing underwater noise. To mitigate the scarcity of jointly annotated underwater data, we construct a high-quality multi-task dataset, FLSea-MTL, from the real-world FLSea monocular depth dataset through foundation-model-guided pseudo-label generation and meticulous manual refinement. Experiments on FLSea-MTL and the MIMIR-UW benchmark demonstrate that MM-MTL achieves state-of-the-art accuracy-efficiency trade-off, with only 8.66M parameters and 10.67G FLOPs, and ablation studies validate the efficacy of the collaborative decoder and the MobileMamba backbone over traditional CNN and Vision Transformer architectures.

We propose Text-Aware Salient Object Ranking (TA-SOR), a multimodal framework for Salient Object Ranking (SOR) that addresses the limitations of purely visual models in capturing semantic dominance and contextual competition among multiple salient instances. TA-SOR integrates visual representations with instance-level text–image attention extracted from a vision–language model, enabling semantic-aware ranking beyond appearance-based cues. An Multimodal-Aware Module performs hierarchical semantic injection through cross-modal interaction, while Semantic-Aligned Cross-modal Attention Guidance and a Ranking Graph Network explicitly model inter-object competition in a unified feature space. To support multimodal learning and evaluation, we construct TASD-ASSR, a semantically aligned extension of the ASSR dataset, and introduce Rb-SOR, a robustness metric based on image utilization rate that complements existing SOR metrics. With approximately 31.7M trainable parameters and frozen visual and language backbones, TA-SOR achieves state-of-the-art performance on ASSR, reaching 0.781 SA-SOR and 0.838 Rb-SOR. The results demonstrate that cross-modal semantic guidance effectively improves ranking accuracy, robustness, and segmentation quality, highlighting the potential of multimodal reasoning for salient object ranking.

Accurate fish volume estimation is essential for sustainable aquaculture management, yet traditional methods are invasive and costly, while existing non-invasive approaches rely on expensive multi-sensor setups. This study proposes a cost-effective infrared (IR)-only pipeline that reconstructs depth and Red Green Blue (RGB) from low-cost infrared videos (3 and coefficient of determination (R²) of 0.961 for volume estimation, outperforming state-of-the-art methods by 19–41% while reducing hardware costs by 80%. This work advances precision aquaculture by providing robust, deployable tools for feed optimization and health monitoring, promoting environmental sustainability amid rising global seafood demand.

The detection and recognition of ships hold significant practical implications for both military and civilian departments. Recent advancements in deep learning technology have led to notable progress in this field. However, the precise detection and recognition of ships remains a challenge, especially in multi-source remote sensing images, due to their different feature expression and resolution. Moreover, the effectiveness of the existing models in complex environments still needs to be further validated. Therefore, in this paper, we construct a new ship dataset which contains 5 distinct ship categories under a number of complex environments. Different from existing datasets that are based on unimodal data, the new dataset uses multi-modal remote sensing images with different resolutions. On this basis, this study introduces an adaptable and robust ship detection and recognition model, namely SIDE-YOLO. The model incorporates a SRCNN and Side Window based contour feature enhancement module (SRSW), a SimAM feature attention Resblock (ResBlockSA), and a DConv based cross-scale feature enhancement block (DCFB), to strengthen the ship edge features, adaptively improve the problem of limited sample in SAR images, and amalgamate multi-scale ship context information, respectively. Validation results demonstrate that the proposed model achieves an average precision rate (mAP) of 84.31%, surpassing state-of-the-art ship recognition models. Notably, for the civilian ship category, the model exhibits 2.17% and 2.63% higher recall and AP in complex scenes, respectively.

To improve the performance of object recognition under artificial prosthetic vision, this study proposes a two-stage method. The first stage is to extract the saliency and edge Mask of the object (SMP, EMP). Then, the irregular visual information of the object is processed using Irregularity Correction (IC). We design eye-hand coordination tasks and simulate artificial vision with retinal prostheses to validate strategy effectiveness, and select direct pixelation (DP) as a control group. Each subject retained a phosphene map in the same stochastic pattern in all his/her trails. The real-time experimental results showed that the deep saliency-based optimization strategies improved the performance of the subjects when completing tasks, in terms of head movement, recognition accuracy, and response time, and counts for successful small-objects recognition. The subjects have the smallest-scale average head movement (76.53 deg ± 20.75 deg), higher average objects recognition accuracy (91.18% ± 2.52%), and less time for finishing the task (35.71 s ± 8.66 s) and better successful search times of the small target objects (1.35 ± 0.33) under the SMP strategy. When integrating with IC, subjects' average performances have further improved to 63.39 ± 15.38 deg, 94.22% ± 3.94%, 25.76 s ± 6.24 s and 1.05 ± 0.30 respectively, which also significantly outperformed the DP condition. These results indicated that when utilizing the deep-learning-based saliency detection and IC processing, subjects could shorten the searching process and were able to discern the target objects more reliably. This work could be informative to future prosthetic devices considering implementation with the technique of artificial intelligence.

Keypoint matching plays a vital role in the realm of synthetic aperture radar (SAR) image processing, serving as a crucial component within this domain. Another key fact to remember keypoint matching is a crucial step in change detection and image stitching. In the context of SAR images, a scale-invariant feature transformation (SIFT)-based approach, known as SAR-SIFT, presents a notable advantage by mitigating the impact of speckle noise.; however, it cannot yield accurate edge information, and the resulting keypoints are nonuniformly distributed. We propose an edge enhancement and homogeneous spatial key point improved SAR–SIFT framework based on an Unbiased Difference-Ratio (UDR) edge detector (called EEKHI–SAR–SIFT) to solve the above problems. The algorithm relies on the characteristics of edge unbiased localization and constant false alarm rate (CFAR) of UDR edge extraction to reduce the extracted wrong corner information as well as enhance the extraction exactness of keypoints. In addition, adaptive non-maximum suppression (ANMS) method is applied to homogenize the dense keypoints with a large initial number that are gained by means of the EEKHI–SAR–SIFT algorithm and reduce their local clustering. Finally, a descriptor construction strategy that retains multi-scale information is adopted to improve the descriptor uniqueness. Tests using multiple sets of SAR image data from different satellites (Gaofen-3, RADARSAT, and Sentinel-1A) demonstrate that the efficacy of the proposed EEKHI-SAR-SIFT algorithm reduces the root mean square error (RMSE) is about 1-2 pixel lower than the final result of the original SAR-SIFT algorithm.