A key limitation of existing learned image compression (LIC) methods is the insufficient utilization of high-level structural priors during bitrate allocation. Without explicit structural guidance, latent quantization is primarily driven by feature statistics, which may lead to suboptimal bit allocation for perceptually important regions such as contours. To address this problem, we propose a contour-guided latent quantization (CGLQ) framework that directly incorporates contour priors into the latent quantization process. Specifically, contour-derived structural cues are transformed into a spatial importance map that modulates the quantization step size, enabling structure-aware and spatially adaptive bitrate allocation. Extensive experiments on the Kodak, CLIC2020, and Tecnick datasets demonstrate the effectiveness of the proposed method. Compared with the widely used Joint’18 baseline, CGLQ achieves consistent improvements across different distortion metrics and reduces BD-rate by up to 30%. Furthermore, the proposed method can be seamlessly integrated into existing LIC architectures and shows competitive performance against recent state-of-the-art approaches, particularly in the medium-to-low bitrate range. These results indicate that explicitly modeling structural priors in the quantization stage is an effective direction for improving perceptual image compression.
Marine biofouling on ship hulls increases hydrodynamic resistance, fuel consumption, and maintenance costs. Although deep learning-based semantic segmentation has been employed for biofouling severity assessment, existing methods still struggle with significant multi-scale morphological variations, complex heterogeneous structures, and severe adhesion among different fouling categories. These challenges often result in inaccurate segmentation, false positives, and missed detections, thereby degrading the reliability of biofouling severity assessment. To address these challenges, this paper proposes an efficient semantic segmentation framework, termed MBSeg. Specifically, a dual-branch Bio Mixer module is designed for collaborative global-local feature modeling by jointly exploiting the global contextual modeling capability of Mamba and the local fine-grained feature extraction capability of CNN. In addition, a lightweight LCBA module based on depthwise separable convolution is introduced to facilitate efficient multi-scale feature refinement and channel alignment. Furthermore, a dedicated pixel-level marine biofouling segmentation dataset is constructed by manually annotating ship hull images that have been previously categorized by domain experts according to biofouling severity, providing a benchmark for model training and future research. Experimental results demonstrate that MBSeg achieves an mIoU of 88.14% on the test set and 75.19% on the challenging test set, while attaining a biofouling severity assessment accuracy of 87.50%. These results demonstrate that MBSeg provides accurate and robust marine biofouling segmentation and offers reliable visual representation for biofouling severity assessment. Code: https://github.com/meng-shengxuan/MBSeg-main
A system for complex scene understanding is presented by integrating multi-modal visual information and conducting multi-element scene analysis. The system consists of two main components: multi-modal perception and multi-element analysis. The perception component integrates an RGB-D camera, an RGB camera, a conventional field-of-view LiDAR, and a wide-angle LiDAR to capture complementary color, depth, and geometric information in complex scenes. The analysis component organizes several established techniques, including simultaneous localization and mapping (SLAM), object detection, human pose estimation, and point cloud processing, within a unified processing framework for scene analysis. To support practical deployment in different spatial conditions, three task-oriented lightweight processing strategies are developed in the system: a wide-angle point-cloud-based space classification method, a crowd-counting and pose-skeleton-based pose analysis method using RGB-D images, and a crowd-counting and person-mask-based pose analysis method using fused image and point-cloud data. Based on these components, the system provides multiple types of scene information, including colorized 3D reconstruction results, space types, unit categories, crowd distribution and count, and human pose categories. The proposed framework emphasizes adaptive processing and system-level integration for multi-modal scene understanding in complex real-world environments. Experimental results in representative indoor and outdoor scenarios demonstrate the feasibility of the overall system and its potential for structured scene representation and human–computer interaction.
