Vous êtes ici : Accueil » Kiosque » Annonce

Identification

Identifiant: 
Mot de passe : 

Mot de passe oublié ?
Détails d'identification oubliés ?

Annonce

19 mars 2018

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


Catégorie : Doctorant


Images acquired by snapshot multispectral cameras

Automatic identification of weed in agriculture fields

Weed control coupled with precision agriculture limits the use of herbicides and is a major challenge for farmers and a priority of the Ecophyto plan. With ‘Chambre d’Agriculture de la Somme’, this PhD thesis will focus on real-time weed identification in images acquired by snapshot multispectral cameras embedded on drones. As these cameras observe outdoor field crops, the lighting and field of view may vary. Therefore, the spatial resolution and spectral properties of images that represent the same weed species may change. 

Texture features from multispectral images acquired under uncontrolled conditions

We propose to first investigate how to adapt the MFSA demosaicing step in any classification scheme of raw texture images acquired by snapshot multispectral cameras. One of the advantages of this approach is to decrease the computation time of the classification process. Then, the PhD student will propose new texture features that are invariant to illumination and spatial resolution changes. Finally he/she will compare the performances reached by the different classification schemes on databases of spectral images that will be provided by ‘Chambre d’Agriculture de la Somme’.

You can find details in PDF

contact: ludovic.macaire@univ-lille1.fr

 

Images acquired by snapshot multispectral cameras

aDigital multispectral cameras sample the visible and near infrared electromagnetic domains into several spectral bands. Such devices provide multispectral images that represent the scene radiance in each spectral band as a separate channel. Multispectral images are used for various application fields such as precision agriculture [1] or waste sorting [2]. The information available in each channel of a multispectral image results from a spectral integration of the product between the scene reflectance, illumination, camera filter transmittances and sensor sensitivity. Therefore, multispectral images depend upon the illumination properties, the camera spectral sensitivity and viewing direction.

Recently snapshot multispectral cameras have emerged to acquire all spectral bands in a single shot. In particular, a single-sensor snapshot technology uses a multispectral filter array (MSFA) to provide a raw image where each pixel is characterized by the value of a single spectral component [3]. To ensure manufacturing practicability, all MSFAs are defined by a basic pattern that respects a trade-off between spatial and spectral sub-samplings. In the following, we consider the non-redundant MSFA formed from the square 5 × 5 basic pattern that is sensitive to 25 spectral bands in the visible and near infrared domains. The 24 component values that miss at each pixel can be estimated by a process known as demosaicing to recover all channels in full definition.

MSFA demosaicing

This process is similar in its principle to the estimation of missing RGB components in raw images captured by single-sensor color cameras fitted with a Bayer color filter array (CFA). CFA demosaicing is a well-studied problem for more than forty years, but MSFA demosaicing is a recent subject with new issues. The principles of spatial and spectral correlations, that exploit the properties of radiance for CFA demosaicing, should indeed be reconsidered.

First, more bands imply a lower spatial sampling rate for each of them, which weakens the assumption of spatial correlation between the raw values of image pixels that sample the same band. Second, since multispectral imaging uses narrow bands whose centers are distributed over the spectral domain, the correlation between channels associated with nearby bands is stronger than between channels associated with distant bands.

The weak spatial and spectral correlations explain why demosaicing schemes are prone to generate strong artifacts in areas with high spatial frequencies. Thus, many demosaicing schemes tend to alter the local high-frequency information that is useful to discriminate the textures. As a result, the demosaiced multispectral image may not well represent the texture observed by the camera, and a texture classification scheme is expected to reach lower performances when applied to a set of demosaiced multispectral images than when applied to the corresponding full-spectral images. 

Texture features from multispectral images

Texture analysis is required in several fundamental problems of image processing like object recognition and texture image segmentation or classification. As the number of labeled spectral images that compose the learning data set is often low whereas the dimension (25) of spectral vectors is high, tools from deep learning do not provide satisfying results and are time consuming [5]. Therefore, a multispectral image is described by some visual cues represented by statistical structures (e.g., histograms), hereafter called texture features. To classify the image, its features are compared with those of reference texture samples.

Texture features that take the spectral information into account improve the characterization of textures, especially when dealing with natural textures observed under uncontrolled lighting conditions (e.g., outdoors). They can be deduced from color texture features. The first and most widely used strategy is the marginal one. It assumes that the texture can be separately described within each channel. Texture features designed for grayscale images are then computed for each channel and aggregated into a global feature representation.

The other main strategy takes advantage of the vector nature of spectral information [6]. Texture features may then provide global statistics about the spatial arrangement of spectral vectors in the image plane. A local analysis may also take spatial correlation between spectral vectors of neighboring pixels into account, and provide texture features like co-occurrence matrices and the histogram of local vector patterns. Multispectral texture characterization by local vector patterns requires to compare spectral vectors of neighboring pixels. As pixels are defined by vectors that are not naturally ordered, we cannot directly compare them.

Automatic identification of weed in agriculture fields

Weed control coupled with precision agriculture limits the use of herbicides and is a major challenge for farmers and a priority of the Ecophyto plan. With ‘Chambre d’Agriculture de la Somme’, this PhD thesis will focus on real-time weed identification in images acquired by snapshot multispectral cameras embedded on drones. As these cameras observe outdoor field crops, the lighting and field of view may vary. Therefore, the spatial resolution and spectral properties of images that represent the same weed species may change.

Texture features from multispectral images acquired under uncontrolled conditions

We propose to first investigate how to adapt the MFSA demosaicing step in any classification scheme of raw texture images acquired by snapshot multispectral cameras. One of the advantages of this approach is to decrease the computation time of the classification process. Then, the PhD student will propose new texture features that are invariant to illumination and spatial resolution changes. Finally he/she will compare the performances reached by the different classification schemes on databases of spectral images that will be provided by ‘Chambre d’Agriculture de la Somme’.

You can find details in PDF

contact: ludovic.macaire@univ-lille1.fr

 

Dans cette rubrique

(c) GdR 720 ISIS - CNRS - 2011-2018.