1 October 2019 in Computer Optics
F“Federal State Unitary Enterprise “State Research Institute of Aviation Systems” (FGUP “GosNIIAS”)
Federal State Unitary Enterprise “State Research Institute of Aviation Systems” (FGUP “GosNIIAS”)
A general approach to a structure-functional analysis and synthesis (SFAS) of deep neural networks (CNN). The new approach allows to define regularly: from which structure-functional elements (SFE) CNNs can be constructed; what are required mathematical properties of an SFE; which combinations of SFEs are valid; what are the possible ways of development and training of deep networks for analysis and recognition of an irregular, heterogeneous data or a data with a complex structure (such as irregular arrays, data of various shapes of various origin, trees, skeletons, graph structures, 2D, 3D, and ND point clouds, triangulated surfaces, analytical data descriptions, etc.) The required set of SFE was defined. Techniques were proposed that solve the problem of structure-functional analysis and synthesis of a CNN using SFEs and rules for their combination.
1 February 2019 in Computer Optics
SSState Research Institute of Aviation Systems
State Research Institute of Aviation Systems
MTMoscow Institute of Physics and Technology
Moscow Institute of Physics and Technology
Face detection is one of the most popular computer vision tasks. There are a lot of face detection approaches proposed including different CNN-based techniques, but the problem of optimal balancing between detection quality and computational speed is still relevant. In this paper we propose new CNN-based solution for face detection called FaceDetectNet. Our CNN architecture is based on ideas of YOLO/DetectNet and GoogleNet architecture supported with some new tools and implementation details created especially for our face detection application. We propose: original iterative proposal clustering (IPC) algorithm for aggregation of output face proposals formed by CNN and the 2-level “weak pyramid” providing better detection quality on the testing sets containing both small and huge images. Our face detection approach is close to previously proposed SSD-based face detection, but the principal difference is that we use the deep features of top hidden CNN layer for forming the face proposals of any size. Thus we utilize the global semantic and context information for improving the detection quality for small faces. Our FaceDetectNet is trained and tested on the most challenging WIDER FACE detection benchmark. Our algorithm achieves the average precision (AP) 0.69 on the WIDER FACE hard level, and thus outperforms all competitive detectors on the Hard level besides the HR state-of-the-art solution. Note that HR solution is based on essentially deeper and slower CNN, while our FaceDetectNet can work in real-time on the NVIDIA GeForce 1080 GPU. On the other hand, SSD-based face detector with comparable CNN parameters provides AP 0.625 only on the WIDER FACE hard level. So, our approach provides the best quality with reasonable computational speed.
4 April 2016 in SPIIRAS Proceedings
VGVladimir Sergeevich Gorbatsevich
Vladimir Sergeevich Gorbatsevich
Приведены результаты исследований по разработке технологий обработки разносенсорной информации, получаемой бортовой системой машинного зрения, для решения комплекса задач навигации и управления беспилотным летательным аппаратом (БЛА). Проведен анализ предметной области, выделены основные задачи, требующие решения для эффективного выполнения основных функций БЛА. Предложены оригинальные методы комплексирования, основанные на диффузной морфологии, разработаны методики подготовки обучающих выборок и глубокого машинного обучения, обеспечивающие высокое качество распознавания, создана база данных синтезированных изображений для обучения алгоритмов распознавания.
1 October 2021 in Computer Optics
S(State Research Institute of Aviation Systems (GosNIIAS)
State Research Institute of Aviation Systems (GosNIIAS)
We introduce attribute and relational representations of mosaic image models with directed relationships between regions. Attribute representations of asymmetric relational models based on stacking, ranking and integral descriptions are considered. We propose some morphological shape similarity measures based on relational models. We show that using the same oriented relational model, various morphological operators can be constructed, in particular, of Serra- or Pyt’ev type. Some constructive methods for the design of such morphological operators in an attribute and relational domains are proposed. From this consideration we also extract a new morophlogical scheme for two-stage mutual adaptive image-and-shape joint filtering: at the first step, the shape is simplified (projected) with regard to the image to be projected, and at the second step, the image is simplified (projected) with regard to the simplified (projected) shape.
1 January 2018 in Computer Optics
MIMoscow Aviation Institute
Moscow Aviation Institute
We propose new morphological conditional estimates of image complexity and information content as well as morphological mutual information. These morphological estimates take into account both the number and the shape of image tessellation (mosaic) regions. We provide such a region shape account via joint use of mosaic image shape models based on the morphological image analysis (MIA) proposed by Yu. Pyt’ev and morphological thickness maps from the mathematical morphology (MM) introduced by J. Serra. Mathematical properties of morphological thickness maps are explored w.r.t. properties of structured elements, and corresponding properties of the proposed morphological image complexity and information content are proved. Some experimental results on image shape comparison in terms of shape complexity and information are reported. Open access images from a Kimia99 database are utilized for these experiments.