Simon C. Roberts

Target detection in SAR imagery by genetic programming

Abstract The automatic detection of ships in low-resolution synthetic aperture radar (SAR) imagery is investigated in this article. The detector design objectives are to maximise detection accuracy across multiple images, to minimise the computational effort during image processing, and to minimise the effort during the design stage. The results of an extensive numerical study show that a novel approach, using genetic programming (GP), successfully evolves detectors which satisfy the earlier objectives. Each detector represents an algebraic formula and thus the principles of detection can be discovered and reused. This is a major advantage over artificial intelligence techniques which use more complicated representations, e.g. neural networks.

Evolving Modules in Genetic Programming by Subtree Encapsulation

In tree-based genetic programming (GP), the most frequent subtrees on later generations are likely to constitute useful partial solutions. This paper investigates the effect of encapsulating such subtrees by representing them as atoms in the terminal set, so that the subtree evaluations can be exploited as terminal data. The encapsulation scheme is compared against a second scheme which depends on random subtree selection. Empirical results show that both schemes improve upon standard GP.

Pragmatic Genetic Programming strategy for the problem of vehicle detection in airborne reconnaissance

A Genetic Programming (GP) method uses multiple runs, data decomposition stages, to evolve a hierarchical set of vehicle detectors for the automated inspection of infrared line scan imagery that has been obtained by a low flying aircraft. The performance on the scheme using two different sets of GP terminals (all are rotationally invariant statistics of pixel data) is compared on 10 images. The discrete Fourier transform set is found to be marginally superior to the simpler statistics set that includes an edge detector. An analysis of detector formulae provides insight on vehicle detection principles. In addition, a promising family of algorithms that take advantage of the GP methods ability to prescribe an advantageous solution architecture is developed as a post-processor. These algorithms selectively reduce false alarms by exploring context, and determine the amount of contextual information that is required for this task.

Evolution of Ship Detectors for Satellite SAR Imagery

A two-stage evolution scheme is proposed to obtain an object-detector for an image analysis task, and is applied to the problem of ship detection by inspection of the SAR images taken by satellites. The scheme: (1) affords practical evolution times, (2) is structured to discover fast automatic detectors, (3) can produce small detectors that shed light into the nature of the detection. Detectors compare favorably in accuracy to those obtained using a SOM neural network.

Evolution of Vehicle Detectors for Infrared Line Scan Imagery

The paper addresses an important and difficult problem of object recognition in poorly constrained environments and with objects having large variability. This research uses genetic programming (GP) to develop automatic object detectors. The task is to detect vehicles in infrared line scan (IRLS) images gathered by low flying aircraft. This is a difficult task due to the diversity of vehicles and the environments in which they can occur, and because images vary with numerous factors including fly-over, temporal and weather characteristics. A novel multi-stage approach is presented which addresses automatic feature detection, automatic object segregation, rotation invariance and generalisation across diverse objects whilst discriminating from a myriad of potential non-objects. The approach does not require imagery to be pre-processed.

The Prediction of Journey Times on Motorways Using Genetic Programming

Considered is the problem of reliably predicting motorway journey times for the purpose of providing accurate information to drivers. This proof of concept experiment investigates:(a) the practicalities of using a Genetic Programming (GP) method to model/forecast motorway journey times, and (b) different ways of obtaining a journey time predictor. Predictions are compared with known times and are also judged against a collection of naive prediction formulae. A journey time formula discovered by GP is analysed to determine its structure, demonstrating that GP can indeed discover compact formulae for different traffic situations and associated insights. GPs felxibility allows it to self-determine the required level of modelling complexity.

Incident Detection on Highways

This chapter discusses the development of the Low-occupancy INcident Detection Algorithm (LINDA) that detects night-time motorway incidents. LINDA is undergoing testing on live data and deployment on the M5, M6 and other motorways in the United Kingdom. It was developed by the authors using Genetic Programming.

Detection of Incidents on Motorways in Low Flow High Speed Conditions by Genetic Programming

Traditional algorithms which set a lower speed limit on a motorway to protect the traffic against collision with a queue are not successful at detecting isolated incidents late at night, in low flow high speed conditions. The Staged Genetic Programming method is used to detect an incident in this traffic regime. The evolutionary engine automatically decides the time duration for the onset of an incident. This method successfully combines traffic readings from the MIDAS system to predict a variety of late night incidents on the M25 motorway.

Textural Classification of Mammographic Parenchymal Patterns with the SONNET Selforganizing Neural Network

In nationwide mammography screening thousands of mammography examinations must be processed. Each consists of two standard views of each breast, and each must be examined by the eyes of an experienced radiologist to determine whether or not to recall the subject, or to undergo a closer mammographic, ultrasound and or more invasive examination. The eyes of the very experienced radiologist are alerted to a mammogram that deserves to be recalled and it is submitted that an ability to pick out anomaly of texture in mammography screening is a very important characteristic of successful outcomes in screening. We digitized a large number of pristine mammography images with a highly accurate scanner and processed textural statistics derived from 450 of these images through a SONNET self-organizing neural network to produce an organization of the mammography archive. In this paper we describe some aspects of this work. The best result produced 39 stable classes and produced relatively narrow classes with an average within-class distance of 1.01 whilst retaining a typical average between-class distance of 2.00. The chief features that discriminated class encodings were the two textural features: angular second moment and contrast. The search for multiscale features over a diverse set of mammograms represents a very challenging problem owing to the high dimensionality of the potential search space.

Initial Exploitation of the SONNET Derived Taxonomy of Mammographic Parenchymal Patterns

A taxonomy of mammography patterns has a number of potential uses which are discussed in this paper. The paper also presents further details about an organization of the mammography archive that was achieved by means of the SONNET self-organizing neural network. Preliminary results on the possible use of the mammography taxonomy to detect cancerous lesions via asymmetry identification are presented. A SONNET hierarchy capable of classifying parenchyma sub-types which combines with evolutionary computation is proposed which may overcome the challenging problem of the search for multiscale features over a diverse set of mammograms.