No 3 (2024)

Image style transfer is an applied task for automatic rendering of the original image (content) in the style of another image (specifying the target style). Traditional image stylization methods provide only a single stylization result. If the user is not satisfied with it due to stylization artifacts, he has to choose a different style. The work proposes a modified stylization algorithm, giving a variety of stylization results, and achieves improved stylization quality by using additional style information from similar styles.

This work presents an approach to reconstructing 3D models from point cloud data, based on the use of modern neural network architectures. The basis of the method is PointNet++ and Transformer. PointNet++ plays a central role, providing efficient feature extraction and encoding of complex 3D scene geometries. This is achieved by recursively applying PointNet++ to nested partitions of the input point set in metric space. Convex decomposition, an important step in the approach, allows transforming complex three-dimensional objects into a set of simpler convex shapes. This simplifies data processing and makes the reconstruction process more manageable. The Transformer then trains the model on these features, allowing for the generation of high-quality reconstructions. It is important to note that the Transformer is used exclusively to determine the position of walls and object boundaries. This combination of technologies allows achieving high accuracy in the reconstruction of 3D models. The main idea of the method is to segment the point cloud into small fragments, which are then restored as polygonal meshes. To restore missing points in point cloud data, a method based on the L1-Median algorithm and local point cloud features is used. This approach can adapt to various geometric structures and correct topological connection errors. The proposed method was compared with several modern approaches and showed its potential in various fields, including architecture, engineering, digitization of cultural heritage, and augmented and mixed reality systems. This underscores its' wide applicability and significant potential for further development and application in various fields.

This paper explores the possibility of using dual geometry representation to improve the speed of ray tracing and ensure the robustness of light propagation simulations in complex optical systems containing free-form surfaces defined by high-order polynomials (up to order 34) or Jacobi polynomials. An analysis was carried out of traditional methods of representing this geometry both in the form of a triangular mesh and in the form of an analytical expression. The analysis demonstrated the disadvantages of traditional approaches, which consist in the insufficient accuracy of calculating the coordinates of the meeting point of the ray with a triangular mesh, as well as the instability of the results of searching for the hit point of tangent rays with the analytically defined surface when using existing calculation methods. As a result, it was proposed to use a dual representation of the geometry in the form of a rough approximation of the surface by a triangular mesh, which is subsequently used as an initial approximation to find the point where the ray hits the surface specified by the analytical expression. This solution made it possible to significantly speed up the convergence of analytical methods and increase their stability. Moreover, using the Intel® Embree library to quickly find the point of intersection of a ray with a coarse triangular mesh and a vector calculation model to refine the coordinates of the point of intersection of the ray with the geometry represented analytically allowed authors to develop and implement a ray tracing algorithm in an optical system containing surfaces with dual geometry representation. Experiments conducted using the developed and implemented algorithm show the significant acceleration of ray tracing while maintaining computational accuracy and high stability of results. The results were demonstrated by calculating the point and flare spread function for two lenses with free-form surfaces defined by Jacobi polynomials. In addition, for these two lenses, the image formed by an RGB-D object simulating a real scene was calculated.

Segmentation of wholeslide histological images through the classification of tissue types of small fragments is an extremely relevant task in digital pathology, necessary for the development of methods for automatic analysis of wholeslide histological images. The extremely large resolution of such images also makes the task of increasing image resolution relevant, which allows storing images at a reduced resolution and increasing it if necessary. Annotating whole slide images by histologists is complex and time-consuming, so it is important to make the most efficient use of the available data, both labeled and unlabeled. In this paper we propose a novel neural network method to simultaneously solve the problems of super-resolution of histological images from 20× optical magnification to 40x and classifying image fragments into tissue types at 20× magnification. The use of a single encoder as well as the proposed neural network training scheme allows to achieve better results on both tasks compared to existing approaches. The PATH-DT-MSU WSS2v2 dataset presented for the first time in this paper was used for training and testing the method. On the test sample, an accuracy value of 0.971 and a balanced accuracy value of 0.916 were achieved in the classification task on 5 tissue types, for the super-resolution task, values of PSNR = 32.26 and SSIM = 0.89 were achieved. The source code of the proposed method is available at: https://github.com/Kukty/WSI_SR_CL.