Chapter 6

CHAPTER 7

Bibliography

7.1. INTRODUCTION

In this thesis we have investigated various image coding standards, methods and algorithms like JPEG, EZW, SPIHT and JPEG 2000, trying to understand and to explain their principles, evaluating their features and comparing them for the purposes. Most of these standards and methods utilise techniques based on the well-known Discrete Wavelet Transform, truly interesting and innovative from the signal, audio, video and still image coding point of view; for this reason we have performed various experiments on the wavelet coefficients to understand better their features within image coding, because there does not exist an exhaustive literature on this subject. To improve the knowledge we decided to look at some tests done by researchers around the world on the usual test images, such as “Lena”, “Barbara” and “Goldhill” for example, but we have deemed more interesting to perform several kinds of tests on images related to the AutoMERS project, aiming to emphasise the features of these images in a way to decide a better coding management.

The first important aspect noticed in these kind of images is the possibility to divide the different parts of the image into :

• Interesting objects, interesting from AutoMERS project point of view, that could be fishes, crabs or other deep sea life forms,
• Un-interesting objects, like the cross used to estimate the other object dimensions, the sinker, the bait or also other fish and life forms already depicted in previous images,
• Background, usually quite uniform, grey or brown, including also the shadows of the other illuminated objects.

• Lossless compression or lossy compression with a low compression ratio and a good visual quality, of the interesting objects within the restored images;
• Lossy compression with very high compression ratio, or no storage or transmission at all, for the other portions of the image; this depends if we are interested in having at least an idea of their features or not.

7.2. CONCLUSIONS ABOUT WAVELET COEFFICIENTS AND IMAGE QUALITY

The various tests with the wavelet coefficients matrices, obtained by applying the DWT to the test images, are done to investigate two features of this transform: the amplitude distribution of the coefficient values and the importance of these coefficients from the point of view of the reconstructed image quality. If we look at the amplitude distribution we can see that a large number of coefficients, about 75-90 %, depending on the image, the QMF used and the detail level chosen, have values lower than 5 % compared to the maximum value; if we compared only the detail level coefficients with the maximum value of the whole matrix we notice that only 1 and 2 percent of them have amplitude values higher than 1/200 and 1/300 of the maximum value respectively. As we will highlight later in this section, the main feature of these kind of matrices is to contain only few high amplitude values, significant for the inverse discrete wavelet transform and for the reconstruction of the image, spread “in a sea” of very low values, considered insignificant. To study the importance of the coefficient amplitudes and their distribution, within the good quality reconstruction of the images, we have decided to see how the image quality, subjective and objective, changes if we consider only a part of these coefficients in the inverse wavelet transform.

A first result is that even if the coefficients of each subband are important for the reconstruction of a good quality image and even if the omission of each of these subbands brings a remarkable loss of quality, the “High-High” subbands of each level contributes in a minor way to the quality maintenance, compared with the “High-Low” and “Low-High” subbands. The most important result is, however, the possibility to go down as far as to use only 1 or 2 percent of the coefficients, the higher amplitude coefficients, to obtain reconstructed images with high values of PSNR, good or more than acceptable visual quality, and low introduction of blur and distortion artefacts. The comparison between several results obtained using different quadrature mirror filters shows that orthogonal filters, Daubechies, Coiflet and Symlet, usually have a lower number of high amplitude coefficients, compared with the biorthogonals; the differences between them are not so high as to influence a choice or a decision, actually for this we have to include also computational complexity, implementation difficulty and other features.

Another two interesting outcomes are obtained by comparing the results derived from experiments performed on different original test images and different image inter-components. Images such as “Ima1”, depicted in Figure 4.5 and showing background areas, contain a lot of quite small edges and changes of luminance; this gives a high number of small wavelet coefficients, significant for the PSNR but insignificant for the AutoMERS point of view. Other images, such as “Ima2” and “Ima5”, contain edges and changes of luminance and colour that are usually fewer but more marked; this results in only a small number of significant coefficients, and gives the possibility to obtain quite high compression ratio with a little and acceptable loss on quality. The possibility to perform inter-component transforms on the RGB images, obtaining the luminance component, Y, and the two colour components, Cb and Cr, is really useful; actually we can compress further the colour components leaving more bits, about 4-8 times more, for the luminance component, obtaining similar PSNR values and visual quality. This coding method greatly improves the compression ratio achieved compared with the coding of the image directly in RGB space, and could be coupled with other methods like 4:2:2 and 4:1:1 downsampling.

7.3. WAVELET TECHNIQUES AND JPEG 2000

The quality and the effectiveness of the wavelet techniques has resulted in the recent development of theseveral wavelet-based image coding algorithms. Comparison of results obtained shows that each of them improves on the old JPEG standard and other methods, both from the quality and compression point of view. Some of them perform better than others, but apart from the EZW coding method, useful for understanding the zero-tree coding, the differences between their PSNR and compression ratio values are not so high as to choose one or the other; for the record we have to say that EQ, CB and SFQ methods give the best results in terms of image quality. The introduction during recent years of a new image coding standard based on wavelet techniques, the JPEG 2000 standard, with its manifold functionality and features, superseded several coding techniques designed solely for improving of compression ratio, to the detriment of other interesting aspects; these aspects are really important for many emerging and future applications such as Internet, medical imaging and network image transmission. With the JPEG 2000 standard, it was intended to create a unified image coding system, providing not only low bit rate operations with improved quality performance. Especially interesting features are embedded bitstream, progressive transmission, scalability, encryption, metadata information, region of interest coding, random code stream access and processing and content based description.

Within this thesis we have tried to explore a little bit in depth both the increment of quality and compression performances, and the possibility to perform ROI coding with excellent results. From the point of view of the image quality and compression ratio, with the help of interesting test results obtained in other research centres, we have developed various experiments with AutoMERS based test images. We can notice that these tests give some astonishing results. Not only does JPEG 2000 improve on JPEG and other coding methods such as SPIHT, increasing the quality of the reconstructed images at the same compression ratio values, with an increment of 2-5 dB of the PSNR value, above all for low bit rate, but also improving the quality of the reconstructed images at the same PSNR values, with sharper images and distortion artefacts less annoying for the human visual system. The JPEG 2000 standard also brings improvements from the point of view of the compression ratio. JPEG could reach compression ratio values of about 30-40 : 1 and 80-90 : 1, maintaining respectively a good and acceptable quality, but JPEG 2000 can reach compression ratio values of about 80-120 : 1 and 200-300 : 1, showing respectively good and acceptable image quality, and can reach 800-1000 : 1 of compression ratio, maintaining a fair representation of the features within the reconstructed images. The lossless compression JPEG 2000 standard is slightly worse than JPEG-LS, an image compression standard created for this purpose, but from the point of view of the compression ratio and the computational complexity, JPEG 2000 performance is really good, giving reconstructed images identical to the original, with a compression ratio of about 4-6 : 1.

One of the most important features of JPEG 2000 for projects like AutoMERS is the possibility to code a region of interest within the image with a higher quality compared with the rest of the image; this feature could be coupled with area or object detection techniques to improve remarkably the compression ratio of the images. Experiments on ROI coding with the test images show that independently from the shape of the ROI, rectangular, circular or arbitrarily shaped, we can achieve a considerable increment of the compression ratio used, maintaining the same quality and PSNR values within the ROI area; actually if the ROI area size is N times smaller than the whole image size, we can obtain a compression ratio N times higher. The results show that there is a linear relation between relative ROI size and bit rate at the same PSNR values; to halve the ROI size means to halve the bit rate, as in Figure 6.27, without changing the quality within the ROI. This feature is good for the project because usually the interesting objects within AutoMERS images have small dimensions, most of the time more than twenty or thirty times smaller than the whole image, and this brings a great saving in transmitted or stored data. For example, from a high definition image of about 28 Mbytes we can store an interesting object (30 times smaller) with an excellent visual quality in a JPEG 2000 file of about 15-20 Kbytes.

7.4. FUTURE DEVELOPMENTS

The possibility to distinguish interesting objects from the rest of the image gives an idea about what kind of improvements could be done on projects like AutoMERS, and which kind of future projects could be developed with the OceanLab and the CEII research centres on this subject. An ideal still image or video camera, useful for AutoMERS project, could be coupled with a system having these features:

• The possibility to identify if new kind of objects are present in the area recorded by the camera; this could be done with a comparison between the new image and images already recorded and stored,
• The possibility to extract objects considered interesting for recording and storage, with techniques of edge detection, object detection and pattern recognition,
• The possibility to compress only these objects, as in the ROI coding, with a good quality or a lossless compression, depending on the purposes, and storing or transmitting them,
• The possibility to classify these objects with the purpose of building a large database; this gives us, besides, the possibility to compare them with database objects, to make some measurements, to add features such as name, date, survey position and other text strings,
• The possibility to zoom into the image, obtaining more detailed images of the interesting objects or selecting a prescribed area of the image, and to zoom out, obtaining panoramic images of the whole area,
• The possibility to suddenly illuminate the area with a white light, or a flash, or if we have a CCD colour camera, to light up with a prescribed light frequency colour so as to not disturb the behaviour of the life forms,
• The possibility for a video camera to change the frame rate adaptively and maybe to record other data such as possible audio traces.

• An improvement of the knowledge of the JPEG 2000 standard and its main feature, ROI coding, with software applications, algorithms and “Ad Hoc” methods, in such a way as to improve the quality, the compression ratio and the computational complexity achievable,
• An investigation into JPEG 2000 Motion Part, in development at this moment, comparing this with other video coding standards, such as object based video coding, MPEG-4, and the context based video coding, MPEG-7, to know if the features introduced by these different standards are useful for the purpose,
• A better knowledge of object detection, pattern recognition and object partition methods, to improve the certainty of retrieving the desired information within the image and to perform estimates and measurements on the recovered objects,
• An investigation into several wavelet based image coding algorithms in recent development, to see if some of these algorithms could be improved with some changes, useful for the purpose, and the possibility of obtaining hardware or software encoders or applications,
• The possibility, from the point of view of zoology studies, to develop software applications that could extract from video or still image data some interesting features such as fish path, fish velocity, life forms behaviour, related to some external events like different light frequency and luminance, even infrared or ultraviolet lights, or sound events, or also electric shocks and magnetic fields,
• Further experiments on wavelet coefficients, matrices, levels and subbands, quadrature mirror filter families, in order to discover other interesting features within AutoMERS-like images,
• Improvements of methods and techniques aimed at improving the quality of the reconstructed images, to remove noise and distortion artefacts introduced and to increase the resolution.