عليرضا ميررشيد

Remote sensing has two main stages: data collection by sensors and digital data transmission to the processing center. Based on the principle of traditional imaging, it can be seen that this process produces a large amount of data and the need for compression before sending to the processing center is necessary. One way to compress images before sending is to use compressed sensing methods. Compressed sensing also has issues such as the need to design a sensing matrix, the need to find the dictionary where the signal is sparse, and time-consuming reconstruction algorithms that make its use difficult in practice. Deep learning can be used as a solution to these problems. The main goal of this research is to use deep learning to perform compressed sensing in remote sensing images. For this purpose, 3 models are presented in this dissertation. The first model in this thesis is in such a way that it has the ability to fully perform compressed sensing. The meaning of complete compressed sensing is the ability to perform sensing by learning the measurement matrix and reconstructing it with end-to-end training of an artificial intelligence model. The design of this model is such that it has improved the reconstruction performance in terms of speed and accuracy compared to the previous models, for example, it has been shown that for a measurement rate of 0.25, the PSNR of the reconstructed image is increased by 1.5dB. By adding the learnable module, the nonlinear quantization operator that is necessary before sending is also trained by the network. In many cases, the main and final goal is not to reconstruct the original image, and high-level decisions about images such as classification and segmentation are considered. Therefore, it seems necessary to provide a solution for making high-level decisions from compressed measured values. This topic is called compressed learning and in this thesis we have presented models to perform two widely used tasks of image segmentation and classification in compressed learning. These models, as the simulation results show, in addition to working with different resolutions of the input image, despite data compression in the network, the performance is comparable to other models in terms of accuracy. For example, it has been shown that in the measurement rate of 0.5, the presented segmentation model has a decrease in the intersection over unio‎n by 0.013dB compared to the modern state-of-the-art models without compression and the presented classification model has a decrease in the classification accuracy by 0.85 compared to the state-of-the-art models without compression, which indicates the appropriate performance of the presented models. Also, both of these models have a better runtime compared to the corresponding state-of-the-art models. For example, it has been shown that the proposed classification model has reduced the response time by more than 30%.

يادگيري عميق , يادگيري فشرده , حسگري فشرده , سنجش از راه دور

Deep learning , Compressed learning , Compressed sensing , Remote sensing

alireza mirrashid

aliasghar beheshti shirazi