Skip to Main content Skip to Navigation
Preprints, Working Papers, ...

Training CNNs on speckled optical dataset for edge detection in SAR images

Abstract : Edge detection in SAR images is a difficult task due to the strong multiplicative noise. Many researches have been dedicated to edge detection in SAR images but very few try to address the most challenging 1-look situations. Motivated by the success of CNNs for the analysis of natural images, we develop a CNN edge detector for 1-look SAR images. We propose to simulate a SAR dataset using the optical dataset BSDS500 to avoid the tedious job of edge labeling, and we propose a framework, a hand-crafted layer followed by learnable layers, to enable the model trained on simulated SAR images to work in real SAR images. The hypothesis behind these two propositions is that both optical and SAR images can be divided into piecewise constant areas and edges are boundaries between two homogeneous areas. The hand-crafted layer, which is defined by a ratio based gradient computation method, helps to tackle the gap between training and testing images, because the gradient distribution will not be influenced by the mean intensity values of homogeneous areas. The gradient computation step is done by Gradient by Ratio (GR) and the learnable layers are identical to those in HED. The proposed edge detector, GRHED, outperforms concurrent approaches in all our simulations especially in two 1-look real SAR images. The source code of GRHED is available at \url{https://github.com/ChenguangTelecom/GRHED}
Complete list of metadatas

Cited literature [38 references]  Display  Hide  Download

https://hal.archives-ouvertes.fr/hal-02424315
Contributor : Chenguang Liu <>
Submitted on : Friday, June 5, 2020 - 10:42:58 AM
Last modification on : Wednesday, June 24, 2020 - 4:19:19 PM

File

main.pdf
Files produced by the author(s)

Identifiers

  • HAL Id : hal-02424315, version 5

Citation

Chenguang Liu, Florence Tupin, Yann Gousseau. Training CNNs on speckled optical dataset for edge detection in SAR images. 2020. ⟨hal-02424315v5⟩

Share

Metrics

Record views

26

Files downloads

15