Papers by Eva Dokladalova
HAL (Le Centre pour la Communication Scientifique Directe), Oct 25, 2021
Recent road crack detection methods obtain appealing scores but typically allow a few pixel toler... more Recent road crack detection methods obtain appealing scores but typically allow a few pixel tolerance margin. This is acceptable for locating cracks, but not for measuring their width (indicator of the cracks' severity). Our baseline model, U-VGG19, obtains an F-score of 71.77% on CrackForest, which is superior to other approaches when no tolerance is admitted. However, increasing the scores without tolerance is difficult due to inaccurate annotations. We propose a novel synthetic dataset, Syncrack, as a benchmark for the evaluation of training with inaccurate annotations. Our results show that inaccurate annotations have a detrimental impact on the F-measure, decreasing it by up to 20%. To overcome this, we study label noise correction techniques using weakly supervised learning. Training U-VGG19 with these corrected labels improves the results on Syncrack by up to 12%. Obtained results on the CrackForest and Aigle-RN datasets support that these approaches are useful for real-life data too.
HAL (Le Centre pour la Communication Scientifique Directe), Jun 26, 2018
Today, drones play an interesting role in the so-called Revolution 4.0. One of the problems studi... more Today, drones play an interesting role in the so-called Revolution 4.0. One of the problems studied by various companies and research groups are the precision landing techniques since this drone feature can be used in applications such as package delivery or object tracking. In this paper, we propose a non-supervised model that allows to detect and recognize a set of landing targets using the Gestalt principles. This proposed method is capable to recognize different coded landing targets in a robust way under outdoor non-controlled light conditions. Comparing to thresholding techniques and other methods, this work deals with image degradations caused by shadows, change of scale, noise and camera target deformation.
HAL (Le Centre pour la Communication Scientifique Directe), Jun 22, 2020
International audienc
HAL (Le Centre pour la Communication Scientifique Directe), Feb 1, 2020
: In this talk we develop three main axes i) design of efficient hardware architectures, ii) comp... more : In this talk we develop three main axes i) design of efficient hardware architectures, ii) computational efficient algorithms targeted for embedded vision systems and iii) hardware support for self-aware computing. We will introduce recent advances within the unifying framework of mathematical morphology. We propose a first morphological processor with arbitrarily large neighborhoods. It allows to obtain previously unachieved performances for serially composed morphological filters, geodesical and conditional operators. The cited processor is based on a novel algorithm formulation of morphological dilation. Finally, the applicative domain will be illustrated in scene understanding context for self aware embedded computing.
Coarse-grained reconfigurable architectures (CGRA) are designed to deliver high-performance compu... more Coarse-grained reconfigurable architectures (CGRA) are designed to deliver high-performance computing while drastically reducing the latency of the computing system. Although they are often highly domain-specifically optimized, they keep several levels of flexibility so that they can be reused. However, their reuse is generally limited due to the complexity of identifying the best allocation of new tasks into the hardware resources. Another limiting point is the complexity to produce a reliable performance analysis for each new implementation. To solve this problem, we propose to consider CGRA as a programmable, configuration-driven computing fabric, called Coarse-Grained Programmable Architecture (CGPA). We propose a new latency-based model to describe all hardware elements. We demonstrate how to implicitly model, with the help of latency’s prediction, the heterogeneity of their material implementations. Our model provides the possibility to assess also the configuration cost, ofte...
2019 IEEE International Conference on Image Processing (ICIP), 2019
Deep convolutional neural networks accuracy is heavily impacted by rotations of the input data. I... more Deep convolutional neural networks accuracy is heavily impacted by rotations of the input data. In this paper, we propose a convolutional predictor that is invariant to rotations in the input. This architecture is capable of predicting the angular orientation without angle-annotated data. Furthermore, the predictor maps continuously the random rotation of the input to a circular space of the prediction. For this purpose, we use the roto-translation properties existing in the Scattering Transform Networks with a series of 3D Convolutions. We validate the results by training with upright and randomly rotated samples. This allows further applications of this work on fields like automatic reorientation of randomly oriented datasets.
Lecture Notes in Computer Science, 2016
Architectural optimization for heterogeneous multi-sensor processing is a real technological chal... more Architectural optimization for heterogeneous multi-sensor processing is a real technological challenge. Most of the vision systems involve only one single color sensor and they do not address the heterogeneous sensors challenge. However, more and more applications require other types of sensor in addition, such as infrared or low-light sensor, so that the vision system could face various luminosity conditions. These heterogeneous sensors could differ in the spectral band, the resolution or even the frame rate. Such sensor variety needs huge computing performance, but embedded systems have stringent area and power constraints. Reconfigurable architecture makes possible flexible computing while respecting the latter constraints. Many reconfigurable architectures for vision application have been proposed in the past. Yet, few of them propose a real dynamic adaptation capability to manage sensor heterogeneity. In this paper, a self-adaptive architecture is proposed to deal with heterogeneous sensors dynamically. This architecture supports on-the-fly sensor switch. Architecture of the system is self-adapted thanks to a system monitor and an adaptation controller. A stream header concept is used to convey sensor information to the self-adaptive architecture. The proposed architecture was implemented in Altera Cyclone V FPGA. In this implementation, adaptation of the architecture consists in Dynamic and Partial Reconfiguration of FPGA. The self-adaptive ability of the architecture has been proved with low resource overhead and an average global adaptation time of 75 ms.
Journal of Real-Time Image Processing, 2015
ABSTRACT This paper focuses on the development of a fully programmable morphological coprocessor ... more ABSTRACT This paper focuses on the development of a fully programmable morphological coprocessor for embedded de-vices. It is a well-known fact that the majority of morpho-logical processing operations are composed of a (potentially large) number of sequential elementary operators. At the same time, the industrial context induces a high demand on robustness and decision liability that makes the appli-cation even more demanding. Recent stationary platforms (PC, GPU, clusters) no more represent a computational bot-tleneck in real-time vision or image processing applications. However, in embedded solutions such applications still hit computational limits. The Morphological Co-Processing Unit (MCPU) replies to this demand. It combines the previously published effi-cient dilation and erosion units with geodesic units to sup-port a larger collection of morphological operations, from a simple dilation to a pattern spectrum by reconstruction. The coprocessor has been integrated into a FPGA plat-form running a server, able to respond client's requests over the ethernet. The experimental performance of the MCPU measured on a wide set of operations brings as results in or-ders of magnitude better than another embedded platform an ARM A9 quad-core processor.
2012 19th IEEE International Conference on Image Processing, 2012
ABSTRACT This paper presents a fast, one-scan algorithm for 1-D morphological opening on 2-D supp... more ABSTRACT This paper presents a fast, one-scan algorithm for 1-D morphological opening on 2-D support. The algorithm is further extended to compute the pattern spectrum during a single image scan. The structuring element (SE) can be oriented under arbitrary angle that makes it possible to perform different orientation-involved image analysis, such as the local angle extraction, directional granulometry, etc. The algorithm processes an image in constant time regardless the SE orientation and size in one scan, with minimal latency and very low memory requirements. For pattern spectra, the C-implementation yields an experimental speed-up of 27× compared to other suitable solutions. Aforementioned properties allow for efficient implementation on hardware platforms such as GPU or FPGA that opens a new opportunity of parallel computation, and consequently, further speed-up.
Journal of Real-Time Image Processing, 2011
... paristech.fr V. Georgiev Faculty of Electrical Engineering, University of West Bohemia, 30614... more ... paristech.fr V. Georgiev Faculty of Electrical Engineering, University of West Bohemia, 30614 Pilsen, Czech Republic e-mail: georg@kae.zcu.cz 123 J Real-Time Image Proc DOI 10.1007/s11554-011-0226-5 Page 2. iterations. The ...
2010 17th IEEE International Conference on Electronics, Circuits and Systems, 2010
This paper describes an original stream implementation of serially composed morphological filters... more This paper describes an original stream implementation of serially composed morphological filters using approximated flat polygons. It strictly respects a sequential data access. Results are obtained with minimal latency while operating within minimal memory space; even for very large neighborhoods. This is interesting for serially composed advanced filters, such as Alternating Sequential Filters or granulometries. We show how the dedicated
2010 IEEE International Conference on Image Processing, 2010
In this paper we present a new pipeline HW architecture for fast 2-D erosions/dilations. The impl... more In this paper we present a new pipeline HW architecture for fast 2-D erosions/dilations. The implementation is based on a recently proposed algorithm allowing to process 2-D data in a stream, minimizing the use of memory and drastically reducing the computing latency. These elementary operators can be chained in an efficient pipeline to realize compound morphological operators (opening, closing, ASF filters, etc.) with no intermediate image storage and minimal latency.
2012 19th IEEE International Conference on Image Processing, 2012
ABSTRACT This paper deals with a dedicated hardware architecture for 1-D morphological opening an... more ABSTRACT This paper deals with a dedicated hardware architecture for 1-D morphological opening and pattern spectrum. These operators allow extraction and measurement of 1-D features in images that is a commonly used technique in image analysis and texture classification. The architecture is based on a recently proposed opening algorithm and makes it possible to obtain arbitrary-oriented opening and granulometry at the same time. Respecting a sequential data access, several instances with different orientation can run in parallel on a single input dataflow, increasing thus the performance (experimentally 414 Mpx/s per opening). It opens applicability of traditionally costly operators in embedded, industrial applications.
Lecture Notes in Computer Science, 2005
Page 1. From Moving Edges to Moving Regions Loic Biancardini1, Eva Dokladalova1, Serge Beucher2, ... more Page 1. From Moving Edges to Moving Regions Loic Biancardini1, Eva Dokladalova1, Serge Beucher2, and Laurent Letellier1 ... 15. Tae Hyeon Kim, Young Shik Moon, A New Flat Zone Filtering Using Morpho-logical Reconstruction Based on the Size and Contrast, VLBV, 1999
Journal of Real-Time Image Processing, 2012
In mathematical morphology, the circular structuring elements (SE) are used whenever one needs an... more In mathematical morphology, the circular structuring elements (SE) are used whenever one needs angular isotropy. Difficult to implement efficiently, the circles are often approximated by convex, symmetric polygons that decompose under the Minkowski addition to 1-D inclined segments.
Lecture Notes in Computer Science, 2011
This paper presents a new streaming algorithm for 1-D morphological opening and closing transform... more This paper presents a new streaming algorithm for 1-D morphological opening and closing transformations on 2-D support. Thanks to a recursive computation technique, the algorithm processes an image in constant time irrespective of the Structuring Element (SE) size, with a minimal latency and very low memory requirements, supporting various input data types. It reads and writes data strictly sequentially in
Lecture Notes in Computer Science, 2008
Spatially variable structuring elements outperform translation-invariant ones by their ability to... more Spatially variable structuring elements outperform translation-invariant ones by their ability to locally adapt to image content. Without restrictions, they suffer from an overwhelming computational complexity. Fast methods for their implementation have recently been proposed for 1-D functions. This paper proposes an extension to 2-D with resizable rectangles.
Uploads
Papers by Eva Dokladalova