Natural image matting is an important problem that widely applied in computer vision and graphics. Recent deep learning matting approaches have made an impressive process in both accuracy and efficiency. However, there are still two fundamental problems remain largely unsolved: 1) accurately separating an object from the image with similar foreground and background color or lots of details; 2) exactly extracting an object with fine structures from complex background. In this paper, we propose an attention transfer network (ATNet) to overcome these challenges. Specifically, we firstly design a feature attention block to effectively distinguish the foreground object from the colorsimilar regions by activating foreground-related features as well as suppressing others. Then, we introduce a scale transfer block to magnify the feature maps without adding extra information. By integrating the above blocks into an attention transfer module, we effectively reduce the artificial content in results and decrease the computational complexity. Besides, we use a perceptual loss to measure the difference between the feature representations of the predictions and the ground-truths. It can further capture the high-frequency details of the image, and consequently, optimize the fine structures of the object. Extensive experiments on two publicly common datasets (i.e., Composition-1k matting dataset, and alphamatting.com dataset) show that the proposed ATNet obtains significant improvements over the previous methods. The source code and compiled models have been made publicly available at https://github.com/ailsaim/ATNet.

Although current scene text detectors achieve remarkable accuracy across different and diverse styles of datasets by fine-tuning models multiple times, these approaches are time-consuming and hinder model generalization. As such, exploring a training strategy that only requires training once on all datasets is a promising solution. However, the text-style mismatch poses challenges to accuracy in such an approach. To mitigate these issues, we propose a style-aware learning network (SLNText) for style-robust text detection in the wild. This includes a style-aware head to distinguish the text styles of images and a dynamic selection head to realize the detection of images with different text styles. SLNText is only trained once, achieving superior performance by automatically learning from multiple text styles and overcoming the style mismatch issue inherent in one-size-fits-all approaches. By using only one set of network parameters, our method significantly reduces the training consumption while maintaining a satisfactory performance on several styles of datasets. Our extensive experiments demonstrate that SLNText achieves satisfactory performance in several styles of datasets, showcasing its effectiveness and efficiency as a promising solution to style-robust scene text detection.

This paper focuses on domain adaptation in a semantic segmentation task. Traditional methods regard the source domain and the target domain as a whole, and the image matching is determined by random seeds, leading to a low degree of consistency matching between domains and interfering with the reduction in the domain gap. Therefore, we designed a two-step, three-level cascaded domain consistency matching strategy—co-occurrence-based consistency matching (COCM)—in which the two steps are: Step 1, in which we design a matching strategy from the perspective of category existence and filter the sub-image set with the highest degree of matching from the image of the whole source domain, and Step 2, in which, from the perspective of spatial existence, we propose a method of measuring the PIOU score to quantitatively evaluate the spatial matching of co-occurring categories in the sub-image set and select the best-matching source image. The three levels mean that in order to improve the importance of low-frequency categories in the matching process, we divide the categories into three levels according to the frequency of co-occurrences between domains; these three levels are the head, middle, and tail levels, and priority is given to matching tail categories. The proposed COCM maximizes the category-level consistency between the domains and has been proven to be effective in reducing the domain gap while being lightweight. The experimental results on general datasets can be compared with those of state-of-the-art (SOTA) methods.

Text detection has been significantly boosted by the development of deep neural networks but most existing methods focus on a single kind of text instance (i.e., overlaid text, layered text, scene text). In this paper, we expand the text detection task from a single dimension to multiple dimensions, thus providing multi-type text descriptions for the scene and content analysis of videos. Specifically, we establish a new task to detect and classify text instances simultaneously, termed TextDC. As far as we know, existing benchmarks cannot meet the requirements of the proposed task. To this end, we collect a large-scale text detection and classification dataset, named Text3C, which is annotated using multilingual labels, location information, and text categories. Together with the collected dataset, we introduce a multi-stage and strict evaluation metric, which penalizes detection approaches for missing text instances, false positive detection, inaccurate location boxes, and error text categories, developing a new benchmark for the proposed TextDC task. In addition, we extend several state-of-the-art detectors by modifying the prediction head to solve the new task. Then, a generalized text detection and classification framework is designed and formulated. Extensive experiments using the updated methods are conducted on the established benchmark to verify the solvability of the proposed task, the challenges of the dataset, and the effectiveness of the solution.

Accurate and robust visual localization under changing environments remains a fundamental challenge in autonomous driving and mobile robotics. Traditional handcrafted features often degrade under long-term illumination and viewpoint variations, while recent CNN-based methods, although more robust, typically rely on coarse semantic cues and remain vulnerable to dynamic objects. In this paper, we propose a fine-grained semantics-guided feature extraction framework that adaptively selects stable keypoints while suppressing dynamic disturbances. A fine-grained semantic refinement module subdivides coarse semantic categories into stability-homogeneous sub-classes, and a dual-attention mechanism enhances local repeatability and semantic consistency. By integrating physical priors with self-supervised clustering, the proposed framework learns discriminative and reliable feature representations. Extensive experiments on the Aachen and RobotCar-Seasons benchmarks demonstrate that the proposed approach achieves state-of-the-art accuracy and robustness while maintaining real-time efficiency, effectively bridging coarse semantic guidance with fine-grained stability estimation. Quantitatively, our method achieves strong localization performance on Aachen (up to 88.1% at night under the $(0.2^{°},0.25 \mathrm{m})$ threshold) and on RobotCar-Seasons (up to 57.2%/28.4% under the same threshold for day/night), demonstrating improved robustness to seasonal and illumination changes.