Browsing by Author "Le Caillec, Jean-Marc"

Now showing 1 - 4 of 4
  • Thumbnail Image
    ItemOpen Access
    Explainability of deep SAR ATR through feature analysis
    (IEEE, 2020-10-20) Belloni, Carole; Aouf, Nabil; Balleri, Alessio; Le Caillec, Jean-Marc; Merlet, Thomas
    Understanding the decision-making process of deep learning networks is a key challenge which has rarely been investigated for Synthetic Aperture Radar (SAR) images. In this paper, a set of new analytical tools is proposed and applied to a Convolutional Neural Network (CNN) handling Automatic Target Recognition (ATR) on two SAR datasets containing military targets.
  • Thumbnail Image
    ItemOpen Access
    Pose-informed deep learning method for SAR ATR
    (The institution of Engineering and Technology (IET), 2020-03-30) Belloni, Carole; Aouf, Nabil; Balleri, Alessio; Le Caillec, Jean-Marc; Merlet, Thomas
    Synthetic aperture radar (SAR) images for automatic target classification (automatic target recognition (ATR)) have attracted significant interest as they can be acquired day and night under a wide range of weather conditions. However, SAR images can be time consuming to analyse, even for experts. ATR can alleviate this burden and deep learning is an attractive solution. A new deep learning Pose-informed architecture solution, that takes into account the impact of target orientation on the SAR image as the scatterers configuration changes, is proposed. The classification is achieved in two stages. First, the orientation of the target is determined using a Hough transform and a convolutional neural network (CNN). Then, classification is achieved with a CNN specifically trained on targets with similar orientations to the target under test. The networks are trained with translation and SAR-specific data augmentation. The proposed Pose-informed deep network architecture was successfully tested on the Military Ground Target Dataset (MGTD) and the Moving and Stationary Target Acquisition and Recognition (MSTAR) datasets. Results show the proposed solution outperformed standard AlexNets on the MGTD, MSTAR extended operating condition (EOC)1, EOC2 and standard operating condition (SOC)10 datasets with a score of 99.13% on the MSTAR SOC10.
  • Thumbnail Image
    ItemOpen Access
    SAR image dataset of military ground targets with multiple poses for ATR
    (SPIE, 2017-10-05) Belloni, Carole; Balleri, Alessio; Aouf, Nabil; Merlet, Thomas; Le Caillec, Jean-Marc
    Automatic Target Recognition (ATR) is the task of automatically detecting and classifying targets. Recognition using Synthetic Aperture Radar (SAR) images is interesting because SAR images can be acquired at night and under any weather conditions, whereas optical sensors operating in the visible band do not have this capability.Existing SAR ATR algorithms have mostly been evaluated using the MSTAR dataset.1 The problem with the MSTAR is that some of the proposed ATR methods have shown good classification performance even when targets were hidden,2 suggesting the presence of a bias in the dataset. Evaluations of SAR ATR techniques arecurrently challenging due to the lack of publicly available data in the SAR domain. In this paper, we present a high resolution SAR dataset consisting of images of a set of ground military target models taken at various aspect angles, The dataset can be used for a fair evaluation and comparison of SAR ATR algorithms. We applied the Inverse Synthetic Aperture Radar (ISAR) technique to echoes from targets rotating on a turntable and illuminated with a stepped frequency waveform. The targets in the database consist of four variants of two 1.7m-long models of T-64 and T-72 tanks. The gun, the turret position and the depression angle are varied to form 26 different sequences of images. The emitted signal spanned the frequency range from 13 GHz to 18 GHz to achieve a bandwidth of 5 GHz sampled with 4001 frequency points. The resolution obtained with respect to the size of the model targets is comparable to typical values obtained using SAR airborne systems. Single polarized images (Horizontal-Horizontal) are generated using the backprojection algorithm.3 A total of 1480 images are produced using a 20° integration angle. The images in the dataset are organized in a suggested training and testing set to facilitate a standard evaluation of SAR ATR algorithms.
  • Thumbnail Image
    ItemOpen Access
    Towards a robust slam framework for resilient AUV navigation
    (Cranfield University, 2021-03) Issartel, Mathieu; Chermak, Lounis P; Le Caillec, Jean-Marc; Richardson, Marrk
    Autonomous Underwater Vehicles (AUVs) are playing an increasing part in modern navies, to the point that the control of oceans will soon be decided by their strategic use. In face of more complex missions occurring in potentially hostile environments, the resilience of such systems becomes critical. In this study, we investigate the following scenario: how does a lone AUV could recover from a temporary breakdown that has created a gap in its measurements, while remaining beneath the surface to avoid detection? It is assumed that the AUV is equipped with an active sonar and is operating in an uncharted area. The vehicle has to rely on itself by recovering its location using a Simultaneous Localization and Mapping (SLAM) algorithm. While SLAM is widely investigated and developed in the case of aerial and terrestrial robotics, the nature of the poorly structured underwater environment dramatically challenges its effectiveness. To address such a complex problem, the usual side scan sonar data association techniques are investigated under a global registration problem while applying robust graph SLAM modelling. In particular, ways to improve the global detection of features from sonar mosaic region patches that react well to the MICR similarity measure are discussed. The main contribution of this study is centered on a novel data processing framework that is able to generate different graph topologies using robust SLAM techniques. One of its advantages is to facilitate the testing of different modelling hypotheses to tackle the data gap following the temporary breakdown and make the most of the limited available information. Several research perspectives related to this framework are discussed. Notably, the possibility to further extend the proposed framework to heterogeneous datasets and the opportunity to accelerate the recovery process by inferring information about the breakdown using machine learning.