Joint Target Detection And Subband Coding For SAR Imagery Data

In recent years, it has become possible to format synthetic aperture radar (SAR) image real time. Future wide-area surveillance system mounted on unmanned air vehicle (UAV) will be capable of collecting SAR imagery at huge coverage rates at a fine resolution. Consideration the large amount of SAR imagery through an available communication link to ground station and to the command centers, image compression techniques are naturally of interest. Therefore, a compression–transmission- decompression strategy is necessary to facilitate real-time transmission. Conventional image compression algorithms are incapable of attaining the required compression while retaining the image fidelity required for processing at the ground station. Since developing a theoretical compression technique isn't a complete solution to the problem, the technique must be realizable in the hardware available for placement on the sensor platform.An algorithm for joint target detection and subband coding for SAR imagery data is investigated in this thesis. The objective is to develop an image compression system which can yield low overall rate while maintaining high quality of target's region of interest. The basic idea behind subband coding is to split up the frequency band of the signal into a number of narrowband, then to code each subband with bit rate accurately matched to the statistics of that band and transmit separately. The main issues in subband coding are design of the analysis and synthesis filters, optimum rate allocation and quantization of subbands. Communication theory on engineering, wavelet theory of mathematics, and image processing on object detection are unified in this dissertation. The main content and contribution are as follow.First, an effective multiresolution target detection algorithm based on 22 subbands decomposition in the transform domain is investigated. The algorithm, without speckle suppression preprocessing, exploits the characteristically distinct variation in speckle pattern with multiresolution model. It is a high parallelizable algorithm with low complexity. With wavelets used in SAR image compression, detection performance is evaluated on simulated and collected SAR imagery data. Db4 wavelet filter is determined to be used in subband coding system.Second, the statistical properties of subbands of various SAR images with different size are studied. Subband coefficients of image data can be modeled using generalized Gaussian distribution (GGD). Both adaptive generalized Gaussian models and fixed generalized Gaussian models are studied for the subbands of SAR images. The estimation method of shape parameter of GGD for different subband is presented, since it plays an important role in the design of quantizer for different subbands. The results show the best value of the shape parameter for all subbands except the LFS is 1.5 and that of the LFS is 2 for the fixed models.Third, in order to make the coding performance insensitive to errors in source modeling, robust quantization is investigated by filtering the input before and after quantization with appropriate filters. All-pass filtering will tend to make a broad class of sources appear Gaussian distribution. A simple way to do this is to scramble the phase spectra of the input source by adding reference phase spectra of the m-sequence with the same size, and then subtracting the same reference phase spectral as postfiltering to recover the original phase spectral of the source. In order to reduce the complex implementation of phase scrambling and descrambling using FFT, only data relevant to ROI is robust quantized.Fourth, the generic ROI mask from inverse db4 wavelet transform is derived so that the low-resolution coding of the background does not adversely affect the high-resolution coding of the target regions. The coding performance of target regions degrades dramatically if the significant subband coefficients are mapped as generic ROI mask. The reason is that significant coefficients expand too fast. To mitigate the situation above, ROI mask is improved with respect to the expression of the inverse db4 wavelet transform, where the items whose coefficients are relatively small are neglected.Fifth, based on trellis coded quantization (TCQ), the method to compress SAR images with an optimal rate allocation strategy is presented. In the proposed system, it is important to encode data in target regions with high bit rate for finer quantization while other data in background with very low bit rate for coarser quantization to achieve large compression. Rate allocation procedure is used to allocate rate in rate-distortion optimization sense for target sequences and background sequences separately. Both fixed rate TCQ (FRTCQ) and entropy constrained TCQ (ECTCQ) are simulated for the proposed system. Experiments on MSTAR datasets show that at low bit rates, SNR of target regions using the proposed system are greater than that using standard coder JPEG2000, context information is also preserved. SNR of target regions obtain from improved ROI mask can be increased by 2 to 6dB. The application performances of the FRTCQ and the ECTCQ based methods are further analyzed.