Cédric Marinel (ongoing)

Modal characterization of mechanical structures by automatic motion analysis from videos

PhD thesis, University of Lille.

Direction: Ludovic Macaire, co-supervision: Benjamin Mathon

Abstract: Vibration monitoring plays a crucial role in ensuring the safety, reliability, and performance of mechanical structures such as bridges and wind turbines. It involves the continuous or periodic measurement and analysis of vibrations to assess the structure state, detect anomalies, and trigger maintenance decisions. Vibration monitoring systems are generally composed of vibration sensors such as accelerometers, a data acquisition device, and a software for data analysis. Because classical contact vibration sensors require complicated setup, this thesis focuses on video-based vibration analysis. Video-based methods perform remote measurements and provide vibration data at each pixel of the video frames. The operational modal analysis from these data determines the structure mechanical properties.
In this doctoral work conducted with EOMYS Engineering company, we investigate video motion estimation methods. We especially focus on phase-based methods that rely on the analysis of the space–frequency decomposition of video frames into complex subbands. These methods provide a dense motion estimation of the scene. A multi-subband approach, based on the phase fusion of the full space-frequency decomposition, is proposed and compared with the
state of the art.
As motion can be estimated at every pixel of each frame, the amount of data is not suited to classical operational modal analysis. Thus, a new video-based operational modal analysis, based on a data reduction technique, is proposed and compared with a state-of-the-art video-based method.
The experimental comparisons are first conducted on synthetic videos of a vibrating cantilever beam, with different motion characteristics. Two experimental setups of straight and bent cantilever beams are finally studied to assess the performances of the methods on real video data acquired by a high-speed camera in controlled conditions.

Anis Amziane

Texture features from multispectral images acquired under uncontrolled conditions. Application to automatic identification of weeds in field crops

PhD thesis, University of Lille, defended on October 18, 2022. Manuscript.

Direction: Ludovic Macaire, co-supervision: Benjamin Mathon

Abstract: Precision spraying aims to fight weeds in crop fields while reducing herbicide use by exclusive weed targeting. Among available imaging technologies, multispectral (multishot) cameras sample the scene radiance according to narrow spectral bands in the visible and/or near infrared domains and provide multispectral radiance images with many spectral channels. The main objective of this work is to develop an automatic recognition system of crop and weed plants in field conditions based on multispectral imaging. We first propose an original multispectral image formation model under the Lambertian surface assumption that takes illumination variation during image acquisition into account. We then propose a method to estimate the reflectance as an illumination-invariant spectral signature, the quality of which is evaluated against state-of-the-art methods and for supervised crop/weed recognition. Because spectral bands associated to the acquired channels may be redundant or contain highly correlated spectral information, we select the best spectral bands to be used in a single-sensor (snapshot) camera model suited for outdoor crop/weed recognition. Finally, we propose an original approach based on a convolutional neural network for spatio–spectral feature extraction from multispectral images at reduced computation costs. Extensive experiments show the contribution of our approach to outdoor crop/weed recognition.

Sofiane Mihoubi

Snapshot multispectral image demosaicing and classification

PhD thesis, University of Lille

Direction: Ludovic Macaire, co-supervision: Benjamin Mathon

Abstract: Multispectral cameras sample the visible and/or the infrared spectrum according to narrow spectral bands. Available technologies include snapshot multispectral cameras equipped with filter arrays that acquire raw images at video rate. Raw images require a demosaicing procedure to estimate a multispectral image with full spatio-spectral definition. In this manuscript we review multispectral demosaicing methods and propose a new one based on the pseudo-panchromatic image estimated directly from the raw image. We highlight the influence of illumination on demosaicing performances, then we propose pre- and post-processing normalization steps that make demosaicing robust to acquisition properties. Experimental results show that our method provides estimated images of better objective quality than classical ones and that normalization steps improve the quality of state-of-the-art demosaicing methods on images acquired under various illuminations.
Multispectral images can be used for texture classification. To perform texture analysis, local binary pattern operators extract texture descriptors from color texture images. We extend these operators to multispectral texture images at the expense of increased memory and computation requirements. We propose to compute texture descriptors directly from raw images, which both avoids the demosaicing step and reduces the descriptor size. For this purpose, we design a local binary pattern operator that jointly extracts the spatial and spectral texture information from a raw image. In order to assess classification on multispectral images we have proposed the first significant multispectral database of close-range textures in the visible and near infrared spectral domains. Extensive experiments on this database show that the proposed descriptor has both reduced computational cost and high discriminating power with regard to classical local binary pattern descriptors applied to demosaiced images.

Multispectral image demosaicing

Masters Thesis, University Lille1 - Sciences et Technologies, defended on August 31, 2015. Manuscript (in French).

Co-supervision: Ludovic Macaire

Arezki Aberkane

CFA image analysis for robust feature extraction.

PhD thesis, University Lille1 - Sciences et Technologies, defended on December 21st, 2017. Manuscript (in French).

Direction: Ludovic Macaire

Abstract: This thesis is devoted to edge detection from the raw image acquired by single-sensor cameras. Such cameras are fitted with a Colour Filter Array (CFA, generally Bayer one) and deliver raw CFA images, in which each pixel is characterized by only one out of the three colour components (red, green, or blue). A demosaicing procedure is necessary to estimate the other two missing colour components at each pixel, so as to obtain a colour image. This however produces artefacts that may affect the performance of low-level processing tasks applied to such estimated images.
We propose to avoid demosaicing to compute the image partial derivatives for edge detection. Simple differentiation kernels, Deriche filters or shifted Deriche filters can be used either in a vector or a scalar approach. The vector approach computes partial derivatives for the three channels and the scalar approach first estimates a luminance image, then computes derivatives. The partial CFA derivatives are then used to compute Di Zenzo gradient for edge detection. We assess the performance of our methods on a large dataset of synthetic images with available edge ground truth. We show that CFA-based approaches may provide as accurate edge detection results as colour vector-based ones at much reduced computation cost.

Yanqin Yang

Objective evaluation of the quality of colour images estimated by demosaicing.

PhD thesis, University Lille1 - Sciences et Technologies, defended on October 8th, 2009. Manuscript (in French).

Direction: Ludovic Macaire

Abstract: Our work deals with the quality of colour images provided by a mono-CCD colour camera, which acquires only one colour component at each pixel by means of the CFA (Colour Filter Array) which covers the CCD sensor. A procedure called demosaicing is necessary to estimate the other two missing colour components at each pixel, so as to obtain a colour image in this kind of cameras. We aim to determine which method of demosaicing provides the results that are best adapted to colour image analyses for the reconstruction of scene. First, we present the principles on how the mono-CCD cameras acquire digital colour images, as well as the different arrangements of CFA used in such cameras. Once the influence of the CFA arrangement on the performance of demosaicing has been presented, we focus our studies on the demosaicing methods based on the Bayer CFA. A mathematical formalization for demosaicing is proposed before we present the numerous demosaicing methods in the literature, as well as the post-processing algorithms to correct the demosaiced images. We then investigate the evaluation criteria for the quality of the colour images estimated by demosaicing. First are described the different possible artefacts generated by demosaicing and the reasons for their generation, which allow us to point out the limits of the classical measures used to evaluate the estimated images. We then propose two original measures to quantify the presence of the two main artefacts, namely false colour and zipper effect. At last, we present new criteria based on the analysis of features extracted from colour images, by measuring the quality of edge detection in the estimated images.

Maxime Devanne

3D modelisation of a human body from Kinect cameras.

Final year project, Télécom Lille1, defended on September 2012. Manuscript (in French).
Co-supervision: Hazem Wannous