Research On Dictionary Learning-based SAR Image Compression Technique

Synthetic Aperture Radar (SAR) is a kind of remote sensing imaging system. SAR images are formed from spatially overlapped radar phase histories and certain signal processing techniques. In a typical SAR data processing and application system, there are many modules related to data transmission and data storage. In the last few years, high-quality images produced by SAR systems, carried on a variety of airborne and spaceborne platforms, have become available. However, while the volume of image data collected by SAR systems is increasing rapidly, the ability to transmit the data at real time, or to store them on the ground station, is not increasing so fast. Thus, there is a strong interest in developing compression methods that can obtain higher compression ratios, while keeping image quality to an acceptable level.In this thesis, sparse representation based model for compression tasks under different scenarios were studied and discussed, including efficient SAR image compression technique based on entropy-constrained dictionary learning, objective fidelity SAR image compression technique based on multi-scale dictionary learning, embedded compression technique based on double-sparsity dictionary learning, and adaptive SAR image compression technique based on the online dictionary learning. The main work includes:(1) Efficient SAR image compression technique based on entropy-constrained dictionary learning. It is mainly used to improve compression efficiency. Sparse representation model mainly concerns about sparse, not directly for the compression. According to the Shannon theorem, the final total number of coding bits equal to the entropy encoding (bits/sample) by the total number of samples. So only considering sparse cannot improve the compression performance. To solve this problem, we have designed a new entropy-constrained dictionary learning algorithm. By making entropy a part of the dictionary learning and sparse coding process, we enhance the SAR image compression efficiency from the root of the signal representation stage.(2) Objective fidelity SAR image compression technique based on multi-scale dictionary learning. It is mainly used to improve the compression efficiency while maintaining small important targets. SAR image reflects the Earth’s surface features, so its image content both exhibit a large number of homogeneous regions, but also contain very fine details target. Maintaining small targets in SAR images is often very important. To solve this problem, we have designed a multi-scale dictionary learning algorithm based on SAR image content quad-tree decomposition and rational bits allocation strategy.(3) Embedded compression technique based on double-sparsity dictionary learning. Embedded coding method refers to the sequence of the compressed bit stream is formed in accordance with the importance of the information they contain. Because SAR images are often used for sensitive tasks (such as military mission), the transmission channel can easily become targets invading party damage, so the embedded coding characteristics of SAR image compression is very important. To solve this problem, we have designed a double-sparsity dictionary learning algorithm for SAR image embedded coding.(4) Adaptive SAR image compression technique based on online dictionary learning. Offline training refers to a fixed dictionary is formed by a certain number of representative training set using a dictionary learning algorithm. However, offline training is lack of generalization ability. To solve this problem, we have designed an online dictionary learning-based compression strategy, making it not only more adaptive for SAR image compression but also suitable for real-time implementation.In summary, this thesis study how to develop SAR image compression techniques under different scenarios, with dictionary learning and sparse representation as the core model. The experimental results showed that the proposed compression methods can solve the corresponding SAR image compression problems under different application scenarios, which can effectively reduce the amount of data transmitted and stored.