Eric Hervet

Comparison of wavelet-based and statistical speckle filters

The wavelet transform has become a very popular tool in signal and image processing. Over the last few years, several authors have proposed wavelet-based filters for speckle reduction in SAR*images, and the results are generally reported to be superior to those obtained with traditional statistical speckle filters. In this paper we give a thorough experimental comparison of representative filters from both categories. We show that spatially adaptive statistical filters yield better noise reduction and preservation of structures than wavelet- based methods, but that the latter have certain advantages compared to statistical filters which are not spatially adaptive.

Automatic birdsong recognition based on autoregressive time-delay neural networks

A template-based technique for automatic recognition of birdsong syllables is presented. This technique combines time delay neural networks (TDNNs) with an autoregressive (AR) version of the backpropagation algorithm in order to improve the accuracy of bird species identification. The proposed neural network structure (AR-TDNN) has the advantage of dealing with a pattern classification of syllable alphabet and also of capturing the temporal structure of birdsong. We choose to carry out trials on song patterns obtained from sixteen species living in New Brunswick province of Canada. The results show that the proposed AR-TDNN system achieves a highly recognition rate compared to the baseline backpropagation-based system

Audio steganalysis using deep belief networks

This paper presents a new steganalysis method that uses a deep belief network (DBN) as a classifier for audio files. It has been tested on three steganographic techniques: StegHide, Hide4PGP and FreqSteg. The results were compared to two other existing robust steganalysis methods based on support vector machines (SVMs) and Gaussian mixture models (GMMs). Afterwards, another classification task aiming at identifying the type of steganographic applied or not to the speech signal was carried out. The results of this four-way classification show that in most cases, the proposed DBN-based steganalysis method gives higher classification rates than the two other steganalysis methods based on SVMs and GMMs.

Retrieval of Images Using Mean-Shift and Gaussian Mixtures Based on Weighted Color Histograms

The topic of this paper is Content-Based Image Retrieval (CBIR) based on colors as a content image descriptor. The tool we developed to that purpose modelizes the colors of an image as a set of 2D Gaussian distributions based on weighted color histograms. Then, given a reference image proposed by a user, the system can automatically classify the image and provide the user with the most similar images to the reference image in its category. Experiments with Corel-1000 dataset demonstrate that our method outperforms the known implementations.

Mapping Aboveground Forest Biomass from Ikonos Satellite Image and Multi-Source Geospatial Data using Neural Networks and a Kriging Interpolation

The present study develops a method for aboveground forest biomass mapping from Ikonos imagery and geospatial data. Reference biomass values by group of species were estimated using Kers equations and inventory data from permanent sample plots (PEP) of 400 m2. A supervised classification of the Ikonos image, based on the maximum likelihood method presenting the five species groups inventoried in the field study, was carried out. Thereafter, various vegetation indices and texture parameters were extracted from the Ikonos image. The extracted Ikonos data were then combined with geospatial data at the same 1 m spatial resolution. Inventory plots biomass values estimated by group of species were used for the neural networks model (Multi-layer Perceptron) training with the backpropagation algorithm. Thereafter, biomass values for sample pixels generated randomly by group of species were predicted with the Multi-layer Perceptron. Then, sample pixels biomass values of each group were used to derive biomass values of other pixels of the same species group by interpolation with the ordinary kriging method using five different variogram models. The Gaussian variogram model yielded the best biomass estimates by comparison with reference biomass values, with percentages of residual errors ranging between 2,6 and 9,8% (absolute value) and percentages of RMSE (root mean square error) ranging between 17.2 and 61.1%.

Speech steganalysis using evolutionary restricted Boltzmann machines

This paper presents a new method to train Restricted Boltzmann Machines (RBMs) using Evolutionary Algorithms (EAs), where RBMs are used in the first step of a steganalysis tool for speech/audio files. The following EAs have been tested: Particle Swarm Optimization (PSO), Genetic Algorithm (GA), Artificial Bees Colony (ABC) and Cat Swarm Optimization (CSO). Our method has been tested with three steganographic techniques: StegHide, Hide4PGP, and FreqSteg. A fourth technique combining the three steganographic methods has also been tested. The results are compared to the conventional contrastive divergence learning algorithm. All EAs outperform the contrastive divergence algorithm.

Mapping forest stands using RADARSAT-2 quad-polarization SAR images: A combination of polarimetric and spatial information

This study proposes an approach which simultaneously uses spatial information and polarimetric data from a RADARSAT-2 quad-polarization satellite image for forest tree species classification. The study area is near the Gounamitz River located in northwestern New Brunswick (Canada). After geometric correction of the image, two statistical models were used for the classification: (1) a Markov random fields model based on an initial segmentation provided by the K-means algorithm to account for the spatial statistical dependencies between adjacent sites; and (2) a K-distribution model with, as parameters, the covariance matrix containing all of the polarimetric information. The classification was optimized using the stochastic simulated annealing algorithm. Validation of the results was performed by comparison with field inventory measurements. Variation of the backscattering coefficient c° obtained for the RADARSAT-2 quad-polarization SAR image with incidence angles of 26 0 and 45 ° ranged from 1 and 3 dB for the different tree species. The results of average and overall accuracies of the classification were respectively 77.13% and 72.35% for the 26° incidence angle image compared to 81.47% and 79.12% for the 45°incidence angle.

Image retrieval based on mean-shift clustering using color descriptor

Content-based indexing of images consists in extracting visual information from digital images (such as pixels, colors, objects, shapes, etc.), and can be performed automatically and fast by computers. This work focuses on color indexing of images. A statistical algorithm called Mean-Shift is used to modelize the color distributions of images as two-dimensional Gaussian kernels. The experiments have been performed on the COIL1 database and the results show that the proposed method compares well to other content-based retrieval methods.

SAR image registration without prefiltering

This article describes a radargrammetric chain for which each step tends to be specific to SAR imagery. First, we start with two complex images acquired in radargrammetric conditions and for which the parameters of the geometric view models are refined. The images are then turned into epipolar geometry while keeping their complex feature. Afterwards a multiresolution method of registration of the two images is used that takes into account a geometric criterion as well as the radiometry specific to SAR images. Finally, a DEM and two orthoimages are computed by spatiotriangulation from the previously found matches and from the geometric view models of the original images.

Applications for deep learning in ecology

A lot of hype has recently been generated around deep learning, a group of artificial intelligence approaches able to break accuracy records in pattern recognition. Over the course of just a few years, deep learning revolutionized several research fields such as bioinformatics or medicine. Yet such a surge of tools and knowledge is still in its infancy in ecology despite the ever-growing size and the complexity of ecological datasets. Here we performed a literature review of deep learning implementations in ecology to identify its benefits in most ecological disciplines, even in applied ecology, up to decision makers and conservationists alike. We also provide guidelines on useful resources and recommendations for ecologists to start adding deep learning to their toolkit. At a time when automatic monitoring of populations and ecosystems generates a vast amount of data that cannot be processed by humans anymore, deep learning could become a necessity in ecology.