Unsupervised methods like deep image prior (DIP) leverage network priors for denoising without labeled data but suffer from slow convergence and overfitting, while deep random projector (DRP) improves efficiency via fixed weights and a random seed yet remains limited by its fully random initialization. In this work, we propose deep fixed projector (DFP), an enhanced DRP-based framework featuring three synergistic improvements: (1) initializing the seed with the noisy image to align optimization with the clean image manifold, (2) using pre-trained clean-to-clean encoder-decoder weights to embed structural priors and accelerate convergence, and (3) introducing inter-inference consistency (IIC), a self-supervised regularization that enforces output stability under input perturbations to suppress noise and reduce overfitting. Experiments show DFP consistently surpasses DIP, DRP, and recent variants in PSNR, SSIM, and LPIPS, while achieving faster convergence and robust denoising quality. Code is available athttps://github.com/Hu-China/Deep-Fixed-Projector.

The deep image prior (DIP) and its variants have demonstrated the ability to address image denoising in an unsupervised manner using only a noisy image as training data, but practical limitations arise due to overfitting in highly overparameterized models and the lack of robustness in the fixed iteration step of early stopping, which fails to adapt to varying noise levels and image contents, thereby affecting denoising effectiveness. In this work, we propose an effective yet fast early stopping metric (ESM) to overcome these limitations when applying DIP models to process synthetic or real noisy images. Specifically, our ESM measures the image quality of the output images generated by the DIP network. We split the output image from each iteration into two sub-images and calculate their distance as an ESM to evaluate image quality. When the ESM stops decreasing over several iterations, we end the training, ensuring near-optimal performance without needing the ground-truth image, thus reducing computational costs and making ESM suitable for application in the denoising of real noisy images.

Point cloud registration serves as a fundamental problem across multiple fields including computer vision, computer graphics, and remote sensing. While local feature descriptors (LFDs) have long been established as a cornerstone for point cloud registration and the LFD-based approach has been extensively studied, the field has witnessed significant advancements in recent years. Despite these developments, the research community lacks a systematic review to consolidate these contributions, leaving many researchers unaware of recent progress in LFD-based registration. To address this gap, we present a comprehensive review that critically examines both state-of-the-art and widely referenced methods across all subtasks of LFD-based registration. Our work provides: (1) an extensive survey of existing methodologies, (2) in-depth analysis of their respective strengths and limitations, (3) insightful observations and practical recommendations, and (4) a thorough summary of relevant applications and publicly available datasets. This systematic overview offers valuable guidance for researchers pursuing future investigations in this domain.

In many applications related to point clouds, registration is an inevitable step when processing point cloud data. The registration methods performed by local shape descriptor (LSD) are computationally efficient and suitable for different scenes, but they achieve low registration accuracy due to the limited performance of the LSD. For this reason, a novel LSD is designed for terrestrial point clouds. First, a simple yet efficient local reference frame (LRF) is developed. The LRF is calculated by the robust normal vector and constant vector, so it has high repeatability. This increases the robustness of the descriptor. Then, the local neighbourhood information is encoded based on the LRF in 3D space. The voxel centers are used to compute the feature descriptor. This increases the descriptiveness of the descriptor because the voxel centers can well preserve the local information. Thus, the proposed LSD is highly descriptive and strongly robust. Based on the novel LSD, a registration method is given. The proposed LSD makes the registration method have high accuracy. Also, the proposed LRF can improve the performance of the correspondence selection, which is an important process in an LSD-based registration method. The experiments performed on the point clouds of different scenes well illustrate that our LRF method has significantly better repeatability and robustness in comparison with other LRF methods. The proposed LRF can largely improve the performance of the descriptor. Our LSD also has significantly better descriptiveness and robustness compared to the other descriptors. As a result, our registration method achieves high accuracy and good time efficiency due to the proposed LSD. The code will be available at: https://github.com/taowuyong?tab=repositories after publication.

In the last decades, 3-D object recognition has received significant attention. Particularly, in the presence of clutter and occlusion, 3-D object recognition is a challenging task. In this article, we present an object recognition pipeline to identify the objects from cluttered scenes. A highly descriptive, robust, and computationally efficient local shape descriptor (LSD) is first designed to establish the correspondences between a model point cloud and a scene point cloud. Then, a clustering method, which utilizes the local reference frames (LRFs) of the keypoints, is proposed to select the correct correspondences. Finally, an index is developed to verify the transformation hypotheses. The experiments are conducted to validate the proposed object recognition method. The experimental results demonstrate that the proposed LSD holds high descriptor matching performance and the clustering method can well group the correct correspondences. The index is also very effective to filter the false transformation hypotheses. All these enhance the recognition performance of our method.

Indoor scene segmentation based on 3-D laser point cloud is important for rebuilding and classification, especially for permanent building structure. However, the existing segmentation methods mainly focus on the large-scale planar structures but ignore the other sharp structures and details, which would cause accuracy degradation in scene reconstruction. To handle this issue, an iterative Gaussian mapping-based segmentation strategy has been proposed in this article, which goes from rough segmentation to refined one iteratively to decompose the indoor scene into detectable point cloud clusters layer by layer. An improved model fitting algorithm based on the maximum likelihood estimation sampling consensus (MLESAC) algorithm is proposed for refined segmentation, which is called the Prior-MLESAC algorithm, to deal with the extraction of both vertical and nonvertical planar and cylindrical structures. The experimental results demonstrate that planar and cylindrical structures are segmented more completely by the proposed strategy, and more details of the indoor structure are restored than other existing methods.