Semi-Automated Magnetic Image Retrieval A J Buckingham

M C Dentith

R D List

The University of Western Australia [email protected]

The University of Western Australia [email protected]

The University of Western Australia [email protected]

SUMMARY Image retrieval systems provide an effective tool for signature mapping and retrieval that can be applied to magnetic images to assist with preliminary interpretation of large datasets. Image retrieval is currently a very active field of research, motivated by the significant increase in the size of digital image databases in a wide range of image-based fields. It has emerged as a powerful tool for searching and locating a desired image, or part-image, from a large image database. Locating discrete circular anomalies sought after when exploring for kimberlites is an example of a potential geophysical application. A model for content-based magnetic image retrieval (CBMIR) using texture and shape descriptors has been developed. Region and boundary-based shape information is extracted using various edge detection techniques, and texture content is derived using statistical and wavelet transform-based methods. The model has been incorporated into a Matlab-based system for image retrieval and results using an experimental magnetic database are presented. The system is interactive, allowing the users intentions to be incorporated into the retrieval results. Tests an the experimental magnetic database, demonstrate that CBIR has the potential to be a powerful tool in magnetic image interpretation, as it has been in other image-based fields. Key words: Aeromagnetic images, automated interpretation, image retrieval, texture analysis, shape analysis

INTRODUCTION The use of images as a data type has increased significantly over the last two decades. Advances in image acquisition, computer processing and storage technologies have allowed the exponential growth of image databases in many fields of science and engineering, including geophysics. As the size of image databases increase, the task of manually searching through a database to retrieve a particular image, or part image, becomes a time consuming, and in some cases, completely unrealistic task. In response to this problem, research in computer vision has focussed on developing ways to automatically extract relevant information from images. The process of searching for, and locating a desired image or part image from a large

image database, is known as image retrieval. There has been much research in the past few years into building efficient image retrieval systems (IRS). Most of the research effort has focused on using low-level image information, which can be automatically derived from an image without any a priori knowledge. This approach is commonly referred to as content-based image retrieval (CBIR). Low-level features including color, shape, texture and spatial relationships are used to characterize the content of an image (Smeulders et al., 2000). Currently there are at least three commercial CBIR systems available and over fifty experimental systems reported both for photographic and domain-specific images. The remote sensing community has embraced CBIR with several domain specific systems, designed, for example, to detect hurricanes in hyper-spectral images (Alber et al., 2001) or to locate volcanoes in SAR data from Venus (Burl et al., 1996). The motivation for applying CBIR technology to magnetic databases is twofold. Firstly, the image is the preferred data type for viewing magnetic data, allowing image-based feature analysis methods to be directly applied to the data. Secondly, image retrieval is a tool well suited to the type of visual signature matching frequently carried out on magnetic data. Locating discrete circular anomalies sought after when exploring for kimberlites is an example of a potential application. The advantage of image retrieval over direct pattern matching techniques is that it allows variability in the response. The system looks for images similar to one or several examples provided, not for images conforming to a simple geometrical shape. A model for content-based magnetic image retrieval (CBMIR) using texture and shape descriptors has been developed. Region and boundary-based shape information is extracted using various edge detection techniques, and texture content is derived using statistical and wavelet transform-based methods. The model has been incorporated into a Matlab-based system for image retrieval and results using an experimental magnetic database are presented. The system is interactive, allowing the users intentions to be incorporated into the retrieval results.

Figure 1 Components of a generalized CBIR system.

ASEG 16th Geophysical Conference and Exhibition, February 2003, Adelaide.

Extended Abstracts

Semi-Automated Magnetic Image Retrieval

CONTENT-BASED IMAGE RETRIEVAL In a typical CBIR application, the user is required to provide an example (called a query) of the type of image being sought. The system then characterizes the query in terms of visual features such as intensity and texture, compares the signature computed to that of each image stored, and returns to the user the most similar images from the database. The stages involved in a generalized CBIR system are illustrated in Figure 1. Depending on the image domain, the database may contain large global images that need to be windowed into overlapping local images. This is the case for satellite and low-level geophysical imagery.

Buckingham, Dentith and List

computer and the high-level concepts (or semantics) used by humans to understand images, CBIR systems are often designed to be interactive (Castelli and Bergman, 2002). The weights assigned to each feature descriptor images may be updated to reflect the degree of relevance assigned to each image by the user. This relevance feedback component improves retrieval accuracy by incorporating the user intentions into the system.

MAGNETIC IMAGE RETRIEVAL Developing a CBIR system for magnetic databases requires careful consideration of the way the data should be represented for analysis. Using a heuristic approach, we have developed a model for content-based magnetic image retrieval (CBMIR) using intensity, texture and shape features. The model is presented in Figure 2. For the purpose of demonstrating CBMIR, a small test database of 14 local images (Figure 3) has been extracted from an aeromagnetic survey flown over a granite-greenstone terrain in the Yilgarn Craton, Western Australia. Each image in the database covers an area of 3.2 km × 3.2 km (64 x 64 pixels). The magnetic signatures in the database are of varying amplitude, directionality and complexity.

Figure 2 A model for CBMIR For each CBIR system, there is a model that details the way image content is characterised and compared (Rui et al., 1998). A set of low-level visual features F = [fi] are used to characterise the input image I. Each feature, such as texture or shape, may be represented in several different ways. For each representation of a given feature rij, a numeric descriptor of the feature dijk is extracted and stored. This descriptor may be a histogram, a matrix of values or a feature vector of discriminating parameters. A similarity measure sij is associated with each feature description, and a weight wij indicating the significance of the particular description is assigned. Quantifying the similarity between a query image and each image in a database is often carried out using a distance measure. For a given query image, the total similarity between the query and each image in the database is calculated. The images are then ranked by their overall similarity to the query presented. The RN most similar images are returned to the user. To deal with the subjectivity of human similarity judgement and the gap between the low-level features extracted by a

Figure 3 The experimental database used to demonstrate CBMIR. Image content is commonly represented (non-numeric) and described (numeric) using intensity based (color and texture) and geometry based (shape) measures. Many methods exist to quantify the intuitive notion of texture. Some CBIR systems use perceptually meaningful measures such as density, regularity and roughness, while others concentrate on first and second order statistics of the image. Considering that texture is a scale-dependant property of an image, the wavelet transform has proven very useful in representing image texture. The energy or mean deviations of the detail coefficients are commonly used to quantify texture. First order statistical measures can be derived from histograms of the wavelet detail coefficients, while co-occurrence matrices computed

ASEG 16th Geophysical Conference and Exhibition, February 2003, Adelaide.

Extended Abstracts

Semi-Automated Magnetic Image Retrieval

Buckingham, Dentith and List

from the detail image coefficients (using an over-complete representation) can describe second order statistics (Wouwer et al., 1999). In many cases, image texture is most prevalent in intermediate or high frequency bands. The wavelet packet transform provides a choice for splitting the frequency bands. A criterion can be used to decide if a sub-image needs to be decomposed further. The adaptive wavelet transform structure of images E and H and presented in Figure 4 (using entropy as the criterion). Wavelet texture features derived from the bi-orthogonal wavelet transform of images E and H (implemented using the lifting scheme of Daubechies and Sweldens, 1998) are also illustrated. The use of shape as a feature for image retrieval has been less developed than texture, mainly due to the inherent complexity of representing it. Shape analysis methods usually require the

Figure 5 Shape analysis of magnetic images For the purpose of evaluating the various image content measures that form the CBMIR model, three geoscientists were given the experimental database and asked to select the two images most similar to each query image. The results are given in Table 1.

Table 1 Similarity results using three human candidates

RESULTS A CBMIR system using the model presented in Figure 2 was written in Matlab and applied to the magnetic database. For selected query images A to N, with RN = 3, the results are presented in Table 2.

Figure 4 Multiscale texture description of magnetic images

representation

and

image to be represented as object regions or boundaries (Loncaric 1998). Depending on the application, this can be achieved through some form of thresholding or edge detection. Separating objects from the background in an image is not a trivial task, and not all types of data are suited to this type of representation. Magnetic images can be represented as closed shapes using the tilt angle approach of Miller and Singh (1994). Magnetisation contrasts (‘magnetic edges’) can be located by tracing the ridges of the horizontal gradient magnitude (HGM) of a reduced to the pole or pseudogravity image. Edge-based shape analysis methods can be applied to the (strength and orientation) image of salient magnetic contrasts. Some shape analysis methods used are illustrated in Figure 4.

Table 2 Retrieval results using the CBMIR system for RN=3 Overall, the automated results correspond reasonably well with the human candidate ‘ground truth’ results. For 9 out of 14 queries, the most similar image retrieved by the CBMIR system was also chosen by one or all human candidates. These results, however, can be improved by making the system interactive. The weights assigned to each feature descriptor can be updated based on the labels assigned to each retrieved image by the user. The approach of Rui et al (1998) has been adopted here.

ASEG 16th Geophysical Conference and Exhibition, February 2003, Adelaide.

Extended Abstracts

Semi-Automated Magnetic Image Retrieval

Buckingham, Dentith and List

refining the model for magnetic images, including perhaps, the use of spatial relationships as a feature.

CONCLUSIONS The application of CBIR to magnetic databases is in its infancy. The appropriate selection of features, feature representations and descriptions requires considerable research. Additionally, if the system is to work effectively on large magnetic databases, features must be extracted rapidly and efficiently. Results using the experimental magnetic database, however, demonstrate that CBIR has the potential to be a powerful tool in magnetic image interpretation, as it has been in other image-based fields.

ACKNOWLEDGEMENTS Figure 6 Retrieval results using human similarity judgement (top), the CBMIR system (middle) and relevance feedback (bottom), for query image A. The labels assigned to the three retrieved images are presented.

I would like to thank the Australian Society of Exploration Geophysics (ASEG) for the research grant assigned to this project, Barrick Gold of Australia for ongoing project support, and Fugro Airborne Surveys for permission to publish the aeromagnetic data.

For query image A, the ground truth results are consistent. The human candidates selected images N and D as being most similar to the query. These results are illustrated in Figure 6. In contrast, the CBMIR system returned images B, J and N as being most similar. Comparing the images returned it is evident that the human candidates placed different emphasis on the visual features used to characterize the images. By allowing the user to label the retrieved images as highly relevant (HR), relevant (R), no opinion (?), nonrelevant (NR) or highly non-relevant (HNR), the weights assigned to each feature descriptor can be updated to reflect individual preference. Results using a single iteration of the relevance feedback component reflect more closely the human candidate response. Both images N and D are returned for RN = 3, as illustrated in Figure 6. Similar improvements can be demonstrated for other query images. The results for query image K are presented in Figure 7. Clearly the human candidates placed importance on the shape content of the image. In contrast, the image judged to be most similar to the query by the system reflects intensity-based similarity. The labels assigned to the retrieved images reflect the users preference for linear crosscutting features. A single iteration of the system using the relevance feedback information produces results consistent with the human response.

DISCUSSION CBIR is a map-based not a model-based approach to image analysis. Retrieval is based purely on the similarity of visual content. This means that for magnetic data, there is no inference to source-body location, geometry or properties. Image retrieval is a tool to assist the interpreter to find magnetic signatures of interest that can then be modeled and manipulated in the usual manner. Features used to characterise image content must be domain specific and highly discriminative. The feature representations and descriptions used for the CBMIR model presented here were selected because they have proven highly effective in other image-based applications. There is however much scope for

Figure 7 Retrieval results using human similarity judgement (top), the CBMIR system (middle) and relevance feedback (bottom), for query image K.

REFERENCES Alber, I. E., Xiong, Z., Yeager, N., Farber, M. and Pottenger, W. M, 2001, Fast Retrieval of Multi- and Hyperspectral Images Using Relevance Feedback. Proceedings of the International Geoscience and Remote Sensing Symposium (IGARSS). Burl, M., Fayyad, U., Perona, P. and Smyth, P., 1996, Trainable cataloging for digital image libraries with applications to volcano detection. California Institute of Technology, Pasadena, CA. Technical Report CNS-TR-96-01. Castelli, V. and Bergman, L. D., 2002, Image Databases: Search and retrieval of digital imagery. John Wiley & Sons.

ASEG 16th Geophysical Conference and Exhibition, February 2003, Adelaide.

Extended Abstracts

Semi-Automated Magnetic Image Retrieval

Buckingham, Dentith and List

Daubechies, I. and Sweldens, W., 1998, Factoring wavelet transforms into lifting steps. J. Fourier Anal. Appl., 4, No. 3, 247 – 269. Loncaric, S., 1998, A survey of shape analysis techniques. Pattern Recognition, 31, No. 8, 983-1001. Miller, H.G. and Singh, V., 1994, Potential field tilt – a new concept for location of potential field sources. Journal of Applied Geophysics, 32, 213-217. Rui, Y., Huang, T. S. and Mehrotra, S., 1998, Relevance Feedback Techniques in Interactive Content-Based Image Retrieval, in Storage and Retrieval for Image and Video Databases, 25-36. Smeulders, A., Worring M., Santini, S., Gupta, A. and Jain, R., 2000, Content-Based Image Retrieval at the End of the Early Years. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22, No. 12, 1349 – 1380. Van de Wouwer, G., Scheunders, P. and Van Dyck, D., 1999, Statistical texture characterization from discrete wavelet representations. IEEE Transactions on Image Processing, 8, No. 4, 592-598.

ASEG 16th Geophysical Conference and Exhibition, February 2003, Adelaide.

Extended Abstracts