Multi-layer insulation, or MLI, is a thermal insulation composed of multiple layers of thin sheets and is often used on spacecraft. It is one of the main items of the spacecraft thermal design, primarily intended to reduce heat loss by thermal radiation. In the case of refilling the satellite in the space, the MLI need first to be cut and then sewed. This procedure is generally done manually which involves manual identification and segmentation.
Our motivation is to automate this process. The reflective and deformable nature of MLI makes segmentation very difficult for traditional computer vision and machine learning techniques. Recent advances in deep learning allow us to approach the problem in a different way. It is widely accepted that Fully Convolutional Network (FCN) has the capability of generating desirable results for image segmentation and we have also shown that it also has the great potential for our application.
FCN first down sample the images using convolution and max pooling and then upsample and classify each pixel with transposed convolution and fully convolutional layers. For MLI detection and segmentation, the fully convolutional layers will output two classes – the background and the object which is the MLI. Prior to train the network, we have captured over 400 images from Robotorium and labeled them to form a dataset. The images capture the MLI in different positions, orientations and lighting conditions, and the labelled imaged outline the contour of the MLI in grayscale.
In our application, FCN32, FCN16 and FCN8 are applied in sequence on the training set, and test the network with different test images which mixed with real images from the Internet. We also include a Conditional Random Field (CRF), in the post-processing field. The result demonstrates that the network can successfully identify and segment MLI in different poses and lightings.
The result for different FCNs and effect of post-processing is shown below.
Comparison of FCNs on subsets
Comparison of results with/without CRF
Codes available at: https://github.com/Jin-Linhao/FCN