Performance evaluation of supervised DL algorithms with domain-specific adapters for object identification of multi-domain satellite images in HPC environments

Abstract

this allows to iteratively achieve effective separation of domain-invariant and -specific features. Second, we investigate deep network training methods to consider various reliability of multiple supervisors. The discussion is made based on an averaging SGD-based approaches, which have been emerged for accelerating the stochastic gradient descent method and ensuring privacy-preservability in the training. As a supervisors’ reliability-aware training strategy, we discuss a method to reduce the impact of label noise in training by exploiting label smoothing. Here, the smoothing rate is determined based on the expected reliability of each supervisor. In conclusion, this dissertation aims to reduce the total processing time required for supervisors' labeling and training deep networks in the case when there exist multiple supervisors with various reliability for multi-domain satellite images, while ensuring the expected reliability above a certain level. Through the performance evaluation, it is shown that the proposed methods not only have high multi-class object identification performance, but also exhibit improved visual explanation. In addition, it is also verified via theoretical models and experiments that they provide the robustness with respect to labeling errors in training the target deep networks.

With the development of remote sensing technology, satellite imagery is now being utilized for a variety of applications such as environmental monitoring. Recently, the need of automating the annotation process is recently emerging

to achieve this, there have been attempts based on convolutional neural networks. Nonetheless, in order for the deep learning-based systems to be applied in practice, there still exists a room for improvement

we focus on the following three points: (a) Since most of analytics applications for satellite imagery are highly critical in their accuracy, wrong decision of the deep learning-based system could cause significant problems. Hence, its explanability on the prediction results becomes extremely important. (b) If dealing with imageries from multiple satellites, we should consider domain shift among them. In this case, the domain shift occurs according to the specification of the image sensors

in deep network training with the multi-domain images, high accuracy should be achieved for all of the source domains. (c) In addition, training data in deep network training for satellite image analytics is mainly obtained from the labeling of supervisors

however, there could exist cases of occurring mislabeling in practice, and the reliability of the supervisors could also be diverse. Hence, it would be important to correct the mislabeling or to perform learning robust to the label noise. To resolve these problems, this dissertation deals with a system for multi-class object identification in a satellite-ground HPC environment

the detailed methods proposed are as follows: First, we propose a deep network architecture for multi-domain learning with domain-specific adapters. The adapter module operates as a plug-in for the backbone network

it has the effect of improving channel and spatial attention for input images as well as extracting domain-specific features. We then present an alternating training strategy of the backbone network and the domain adapters

the training process is conducted by alternatively freezing the two components