B.R. Calder

Spatial stochastic models for seabed object detection

We introduce two statistical models designed to detect discrete objects in sidescan SONAR which consider complimentary approaches to the problem. The first considers a complex textural model for the objects and implements detection through a dual hypothesis on texture class presence, while the second implements a complex Gibbs field model of the image and utilizes prior knowledge of typical object morphologies to support its detection rate. The models are demonstrated on examples of different seabed sediments and object types, and are shown to be reliable in operation. The common theme of the two models is use of spatial context in analysis, which, we argue, is a very powerful technique for improving the flexibility and reliability of any analysis system.

Bayesian interface detection in very shallow chirp seismic data

In this paper we consider an approach to the problems of processing very high resolution, very shallow, seismic data. We have developed a processing strategy based on a Bayesian model of the basebanded, matched filtered, signal. We have found this model to be robust in detecting close reflector wavelets (overlapping by up to 80%) and in adapting to local conditions within the data under suitable stochastic a priori constraints. In addition, the use of Reversible-Jump Markov chain Monte Carlo techniques allow us to address the issue of model selection directly. After developing the requirements for the model, and describing the processing methodology, we show results in synthetic and real data sets. We show that under realistic operational conditions, the algorithm is capable of resolving subtle layers, making subsequent interpretation simpler.

Constrained image restoration with a multinomial prior

A Bayesian image processing model is proposed based on, a Markovian multinomial prior. The technique has application in texture segmentation where its introduction of spatial context can improve segmentation accuracy by 60%. Other applications include general image restoration where 18 dB SNR improvement is possible. In addition, the computational complexity of the system is low, making it ideal as a component part of other systems. We show quantitative experiments to illustrate the performance of the algorithm, and groundtruth examples are provided to show the effect in practice.