به سمت شبکه های عصبی پیچشی عمیق با منبع مقرون به صرفه
Abstract
1- Introduction
2- Related literature
3- Quantized convolutional neural networks for HSI segmentation
4- Experiments
5- Conclusions and future work
References
Abstract
Hyperspectral image analysis has been gaining research attention thanks to the current advances in sensor design which have made acquiring such imagery much more affordable. Although there exist various approaches for segmenting hyperspectral images, deep learning has become the mainstream. However, such large-capacity learners are characterized by significant memory footprints. This is a serious obstacle in employing deep neural networks on board a satellite for Earth observation. In this paper, we introduce resource-frugal quantized convolutional neural networks, and greatly reduce their size without adversely affecting the classification capability. Our experiments performed over two hyperspectral benchmarks showed that the quantization process can be seamlessly applied during the training, and it leads to much smaller and still well-generalizing deep models.
Introduction
Hyperspectral imaging (HSI) is being continuously applied in a variety of fields, including biochemisty, biology, mineralogy, and remote sensing [90]. It captures a wide spectrum of light, and forms an array of usually more than a hundred of reflectance values acquired for every pixel in the image. Such amount of information can be effectively used to classify each pixel to a specific class, and to find the boundaries of objects within a scene imaged using a hyperspectral sensor in the process of HSI segmentation1 [62]. The remote sensing community currently struggles with applying hyperspectral segmentation engines in constrained hardware settings, in the context of on-board Earth observation. It is in contrast to the post processing of such imagery which is performed back on Earth, after transferring images from a satellite equipped with a hyperspectral camera. This data transfer is extremely costly and time-consuming, and it is not feasible in the majority of Earth observation use cases where short re-visit times that can be seen as the temporal resolution of hyperspectral data, and rapid response to the events captured within a scene are critical practical issues. Disaster prevention, monitoring, and post-crisis operation alongside precision agriculture are the most notable examples of such applications [80].