Research On Sonar Image Quality Assessment For Self-decision Compression

Because of the ability to take images in relatively dim light underwater,sonar has been more and more widely used in underwater rescue,ocean exploration,seabed mapping,etc.,during recent years.In practice,sonar images can be analyzed and processed after being transmitted via underwater acoustic channel.It is necessary to compress data to meet the need of extremely limited underwater channel bandwidth,especially for video and image transmission with high volumes of data.Since human visual system is inclusive to the distortion of multimedia data,optimizing system according to Quality of Service(QoS)like bit error ratio refers to enormous stress on link quality control.On the other hand,underwater acoustic channel is time-varying,but conventional system only provides with one fixed scheme.In order to achieve adaptive sonar image compression,which adapts to the changing condition of underwater acoustic channel,the selection of compression parameter,compression scheme and retansmission scheme is guided by the quality of compressed image and received image in this paper.There is few related work on sonar image quality evaluation at present.A standard sonar image quality database is first built in this paper.Then the characteristics and applications of sonar images are discussed based on this database.Finally,three sonar image quality assessment methods(full-reference,reduced-reference and no-reference)are proposed in this paper to evaluate the qualities of compressed sonar images and received sonar images.To tackle the problem of sonar image quality assessment,the applications of sonar images and human visual system are both taken into account.When the image has relative high quality,perceived information receives great interest.However when a sonar image is afflicted with sever degradations,people tend to pay attention to image components like edge and texture.Unlike common-used global entropy,the local entropy,which gives consideration to both the interlay between adjacent pixels and"uncrowded window",is extracted as global information.The edge map,which is important to sonar image applications,is employed in this paper to represent structural information.Based on the saliency-based pooling,the changes of statistical and structural information are pooled into two parameters.Finally,we map the proposed features to the subjective image quality by building a quadratic polynomial model,which avoids severe overfitting caused by complex models.Most sonar images are used for practical applications,image understanding is one of the main factors affecting subjective quality of sonar images.Considering the contributory factor and determinant factor of image understanding,first,the similarity of the normalized histogram of block edge map between the original and distorted sonar images is employed in this stage to represent the contour information of a sonar image.Then the image entropy,kurtosis and skewness are extracted to represent the influence of information and comfort indices to image understanding.Finally,the information and comfort indices and the structural similarity are systematically integrated for quality prediction.One of the characteristics differ sonar images from natural scene images,is the importance of macroscopic and microcosmic structures.For optical natural scene images,the macroscopic and microcosmic structures share the same importance,while for sonar images,macroscopic structures occupy the main status.Image quality assessment methods for optical natural scene images often do not distinguish between macroscopic and microcosmic structure.In this paper,a series of sparsity-based features descriptive of contour information are firstly extracted from frequency and spatial domains.The contour degradation degree is then measured by calculating the ratios of extracted features before and after filtering.It has been verified in this paper that the contour degradation degree is relevant to the distortion contained in the sonar image.Finally,a bootstrap aggregating(bagging)based support vector regression(SVR)module,which avoids overfitting,is trained to build the relationship between contour degradation degree and SI quality.