Deep background subtraction with scene-specific convolutional neural networks

Braham, Marc; Droogenbroeck, Marc Van

doi:10.1109/iwssip.2016.7502717

Cited by 264 publications

(224 citation statements)

References 20 publications

(31 reference statements)

Supporting

Mentioning

224

Contrasting

Order By: Relevance

“…Once our model is trained on this labeled dataset, it can be applied to the input ADI sequence for evaluation without risk of overfitting 3 . The fact that SODIRF and SODINN can be trained on a labeled dataset created from a given ADI sequence means that these models are fine-tuned to each ADI sequence (Braham & Van Droogenbroeck 2016). We have tested SODIRF and SODINN on coronagraphic ADI sequences from different instruments.…”

Section: From Unsupervised To Supervised Learningmentioning

confidence: 99%

Supervised detection of exoplanets in high-contrast imaging sequences

González

Absil

Droogenbroeck

2018

A&A

View full text Add to dashboard Cite

Context. Post-processing algorithms play a key role in pushing the detection limits of high-contrast imaging (HCI) instruments. State-of-the-art image processing approaches for HCI enable the production of science-ready images relying on unsupervised learning techniques, such as low-rank approximations, for generating a model PSF and subtracting the residual starlight and speckle noise. Aims. In order to maximize the detection rate of HCI instruments and survey campaigns, advanced algorithms with higher sensitivities to faint companions are needed, especially for the speckle-dominated innermost region of the images. Methods. We propose a reformulation of the exoplanet detection task (for ADI sequences) that builds on well-established machine learning techniques to take HCI post-processing from an unsupervised to a supervised learning context. In this new framework, we present algorithmic solutions using two different discriminative models: SODIRF (random forests) and SODINN (neural networks). We test these algorithms on real ADI datasets from VLT/NACO and VLT/SPHERE HCI instruments. We then assess their performances by injecting fake companions and using receiver operating characteristic analysis. This is done in comparison with state-of-the-art ADI algorithms, such as ADI principal component analysis (ADI-PCA).Results. This study shows the improved sensitivity vs specificity trade-off of the proposed supervised detection approach. At the diffraction limit, SODINN improves the true positive rate by a factor ranging from ∼2 to ∼10 (depending on the dataset and angular separation) with respect to ADI-PCA when working at the same false positive level. Conclusions. The proposed supervised detection framework outperforms state-of-the-art techniques in the task of discriminating planet signal from speckles. In addition, it offers the possibility of re-processing existing HCI databases to maximize their scientific return and potentially improve the demographics of directly imaged exoplanets.

show abstract

Section: From Unsupervised To Supervised Learningmentioning

confidence: 99%

Supervised detection of exoplanets in high-contrast imaging sequences

González

Absil

Droogenbroeck

2018

A&A

View full text Add to dashboard Cite

show abstract

“…Other authors emulated background subtraction strategies based on DNNs framework [10][11][12]23]. DNNs exhibit excellent performances in segmentation tasks such as semantic segmentation tasks [22,24,25] and foreground segmentation tasks [26].…”

Section: Related Workmentioning

confidence: 99%

“…Some researchers have begun using deep neural networks (DNNs) for background subtraction [8][9][10][11][12]. Zhang et al [8] and Shafiee et al [9] used high-dimensional features from DNNs for background modeling.…”

Section: Introductionmentioning

confidence: 99%

“…Zhang et al [8] and Shafiee et al [9] used high-dimensional features from DNNs for background modeling. Some authors [10][11][12] proposed a background subtraction framework using a convolutional neural network. DNN-based methods automatically discover the background features from training images.…”

Section: Introductionmentioning

confidence: 99%

“…In our previous work [13], we investigated the fundamental behaviors of the DNN based on Braham's network [10]. Therein, we mainly focused on the first and last layer.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Analytics of Deep Neural Network-Based Background Subtraction

et al. 2018

View full text Add to dashboard Cite

Deep neural network-based (DNN-based) background subtraction has demonstrated excellent performance for moving object detection. The DNN-based background subtraction automatically learns the background features from training images and outperforms conventional background modeling based on handcraft features. However, previous works fail to detail why DNNs work well for change detection. This discussion helps to understand the potential of DNNs in background subtraction and to improve DNNs. In this paper, we observe feature maps in all layers of a DNN used in our investigation directly. The DNN provides feature maps with the same resolution as that of the input image. These feature maps help to analyze DNN behaviors because feature maps and the input image can be simultaneously compared. Furthermore, we analyzed important filters for the detection accuracy by removing specific filters from the trained DNN. From the experiments, we found that the DNN consists of subtraction operations in convolutional layers and thresholding operations in bias layers and scene-specific filters are generated to suppress false positives from dynamic backgrounds. In addition, we discuss the characteristics and issues of the DNN based on our observation.

show abstract