Sonar Image Recognition And Super-Resolution Reconstruction Based On Sparse Representation

Sonar image recognition and super-resolution reconstruction are two crucial technologiesin future shipping and ocean industry. Whose developments are of great importance in civilianuse and military affairs.In recent years, sparse representation gains much attention ofresearchers, and it has been applied to image compression, image de-noising and imagerestoration. Considering the complexed underwater envioment and the characteristics of sonarimage, sonar image recognition and super-resolution reconstruction are studied based onspares representation in this thesis. Main study contents are as follows:The thesis summarized the development of image recognition, super-resolutionreconstruction and sparse representation. Sparse representation is introduced in the process ofsonar image recognition and super-resolution reconstruction. Three important parts ofcompression sensing method in sparse representation are studied, which are sparse-basis,observing matrix and reconstruction algorithm. Sonar image recognition has great noises,during the preprocessing of sonar image, a appropriate de-noising method is used to removethe noises of sonar image, and then its normalized treatment is given.Considering the local feature’s importance of forward looking sonar image,Non-negative Matrix Factorization is introduced. Because compressed sensing must meet theneed of irrelevance between measurement matrix and sparse-base matrix, so Non-negativeMatrix Factorization is improved. Compressed sensing recognition method is proposed by thecombination of improved NMF and sparse representation. Simulator show that the proposedmethod has great recognition rate effect.The forward looking image is often influenced by bubbles, vesicles and occludes, whichdecreases the recognition effect quickly. Sparse representation can eliminate the occlusive partefficiently through the method of adding occlusive dictionary. But after characteristicsextraction, the atom number are huge, which increases the recognition time. According to theabove problems, a new occlusive dictionary is designed based on dictionary learning, whichdecreases the atom number of the dictionary, increasethe real-time performance of the sonarimage recognition, and obtains same recognition rate compared with original occlusivedictionary.Aiming at the important texture information and rotation invariance of side scan sonarimages, gray level-gradient co-occurrence matrix feature extraction method is introduced,which is used to compare two kinds of gradient-solving effect between Roberts operator and Sobel operator. After gray level-gradient co-occurrence matrix extraction, the texture numbercharacteristics can replace the gray level information characteristics to recognize the sparserepresentation as sparse basis matrix, which leads to good recognition effect and rotationinvariance.For sonar image, the noise influence is great, the expression of the smooth componentcan’t be neglected. Sonar image contains smooth, edge, and texture parts. Using discretestable wavelet basis, Contourlet basis and Gabor basis to build a dictionary to depict the threeparts of sonar image, multi frame sonar image super-resolution reconstruction model based onmulti-layer sparse representation is constructed. Multi frame sonar image are beneficial to thecomplement of sonar image in formations, and extended Basis Pursuit De-noisingreconstruction algorithm is proposed, which obtains good super-resolution reconstructioneffect and robustness.