John A. Marchant

Shadow-invariant classification for scenes illuminated by daylight.

A physics-based method for shadow compensation in scenes illuminated by daylight is proposed. If the daylight is represented by a simplified form of the blackbody law and the camera filters are of infinitely narrow bandwidth, the relationship between red/blue (rm) and green/blue (gm) ratios as the blackbodys temperature changes is a simple power law where the exponent is independent of the surface reflectivity. When the CIE daylight model is used instead of the blackbody and finite bandwidths for the camera are assumed, it is shown that the power law still holds with a slight change to the exponent. This means that images can be transformed into a map of rm/gmA and then thresholded to yield a shadow-independent classification. Exponent A can be precalculated from the CIE daylight model and the camera filter characteristics. Results are shown for four outdoor images that contain sunny and shadowed parts with vegetation and background. It is shown that the gray-level distributions of the pixels in the transformed images are quite similar for a given component whether or not it is in shadow. The transformation leads to bimodal histograms from which thresholds can easily be selected to give good classifications.

Comparison of a Bayesian classifier with a multilayer feed-forward neural network using the example of plant/weed/soil discrimination

The feed-forward neural network has become popular as a classification method in agricultural engineering as well as in other applications. This is despite the fact that statistically based alternatives have been in existence for a considerable time. This paper compares a Bayesian classifier with a multilayer feed-forward neural network in a task from the area of discriminating plants, weeds, and soil in colour images. The principles behind and the practical implementation of Bayesian classifiers and neural networks are discussed as are the advantages and problems of each. Experimental tests are conducted using the same set of training and test data for each classifier. Because the Bayesian classifier is optimal in the sense of total misclassification error, it should outperform the neural network. It is shown that this is generally the case. There are significant similarities in the performance of each classifier. Understanding why this should be the case gives insight into the operation of each classifier and so the paper explores this aspect. In this work, the Bayesian classifier is implemented as a look-up table. Thus any probability function can be represented and the decision surfaces can be of any shape, i.e. the classifier is not restricted to a linear form. On the other hand, it does require a relatively large amount of memory. However, memory requirement is no longer such a major issue in modern computing. Thus, it is concluded that if the number of features is small enough to require a feasible amount of storage, a Bayesian classifier is preferred over a feed-forward neural network.

Physics-based colour image segmentation for scenes containing vegetation and soil

Colour segmentation of images containing vegetation and soil is the theme of this work. Physics-based reflection models are used to develop an algorithm for separating object pixel clusters in the three-dimensional red, green and blue colour space. The dichromatic reflection model that is used as the basis for this algorithm, defines a plane in which the pixels from an object of a given colour will lie. The illuminant colour and the intrinsic body colour of the object determine the parameters of the dichromatic plane. Scenes containing objects that differ in colour, form multiple dichromatic planes in RGB space but the illuminant vector is common to all planes. The algorithm therefore counts the number of image pixels that intersect with a plane formed by the illuminant vector and its normal as the plane is rotated around the illuminant vector. In images comprising two objects that differ in colour, vegetation and soil for example, the method produces a bimodal histogram where the two modes correspond to clusters of pixels from the two objects. Data on plant reflectance spectra can be used to identify which of the clusters is vegetation. The performance of the method is assessed using receiver operator characteristic curves and the probability of misclassification is measured.

Accurate blemish detection with active contour models

Abstract This paper presents a novel image analysis scheme for accurate detection of fruit blemishes. The detection procedure consists of two steps: initial segmentation and refinement. In the first step, blemishes are coarsely segmented out with a flooding algorithm and in the second step an active contour model, i.e. a snake algorithm, is applied to refine the segmentation so that the localization and size accuracy of detected blemishes is improved. The concept and the formulation of the snake algorithm are briefly introduced and then the refinement procedure is described. The initial tests for sample apple images have shown very promising results.

Evaluation of an imaging sensor for detecting vegetation using different waveband combinations

Abstract This paper uses data collected from an earlier reported imaging sensor to investigate the classification of vegetation from background. The sensor uses three wavebands, red; green; and near infra-red (NIR). A classification method (the alpha method) is introduced which is based on a model of the light source and the reflecting surface. The alpha method is compared with two ratio methods of classification (red/NIR and red/green) and two single waveband methods of classification (NIR and green intensity). The Receiver Operating Characteristic Curve (ROC) is used to evaluate the classifications on realistic test images. ROCs compare the ‘true positive ratio’ with ‘the false positive ratio’ as the classification parameter varies. The area under the ROC gives a measure of how well an algorithm performs. Measurements on the ROC show that the alpha and ratio methods all perform reasonably well with the red/green ratio giving slightly poorer performance than the alpha method and the red/NIR ratio. The single waveband methods perform significantly less well with green intensity easily the worst. The alpha and ratio methods have ‘best’ thresholds that correspond with detectable histogram features when there is a significant amount of vegetation in the image. The physical basis for the alpha method means that there is a detectable mode in the histogram that corresponds with the ‘best’ threshold even when there is only a small amount of vegetation. The single waveband methods do not produce histograms, which can easily be analysed, and so their use should be confined to simple images.

Spectral invariance under daylight illumination changes

We develop a method for calculating invariant spectra of light reflected from surfaces under changing daylight illumination conditions. A necessary part of the method is representing the illuminant in a suitable form. We represent daylight by a function E(lambda, T) = h(lambda)exp[u(lambda)f(T)], where lambda is the wavelength, T is the color temperature, h(lambda) and u(lambda) are any functions of lambda but not T, and f(T) is any function of T but not lambda. We use an eigenvalue decomposition on the logarithm of the CIE daylight standard at various color temperatures to obtain the necessary functions and show that this gives an extremely good fit to CIE daylight over our experimental range. We obtain experimental data over the range 350-830 nm from a range of standard colored surfaces for 50 daylight conditions covering a wide range of illumination spectra. Despite a considerable variation in the spectra of the reflected light, we show only small variations when the transformation is used. We investigate the possible causes of the residual variation and conclude that using the above approximation to daylight is unlikely to be a major cause. Some variation is caused by local daylight conditions being different from the CIE standard and the rest by measurement and modeling errors.

Testing a measure of image quality for acquisition control

Abstract Previous work by the author has shown that the entropy of an images histogram can be used to control the acquisition variables (brightness, contrast, shutter speed) of a camera/digitiser combination in situations where the imaging conditions are changing. Although the control leads to histograms that satisfy pragmatic expectations of what a ‘good’ histogram should look like (i.e. filling the dynamic range of the digitiser without too much saturation), it avoids the problem of what we mean by a good histogram in the machine vision context and whether the control produces images that have these histograms. In this work a good image is defined to be one where the subsequent analysis algorithms work well. Three different algorithms, each containing many diverse components, are tested on sets of images with different acquisition parameters. As well as acquiring at different parameters, a simulation of the image acquisition process is derived and validated to assist evaluation. Test results show that near-optimal performance is obtained with maximum entropy and it is concluded that this measure is a suitable one for control of image acquisition.

Color invariant for daylight changes: relaxing the constraints on illuminants

We extend previous work that addressed the problem of color changes on reflective surfaces resulting from changes in the daylight spectrum. In that work, we constrained the illuminants to a family represented by the Wien approximation to Plancks formula in order to derive a function of the three camera outputs that is invariant to daylight changes. In this work, we show that the constraint on the form of the illuminants can be relaxed and that a much more general form is permissible. We use principal components analysis on the logarithm of the illumination to represent the CIE standard in the more general form and show that the result closely represents the standard. We recalculate the exponent used in the invariant for our camera from the extended theory and obtain a result that duplicates the one found by empirical means used in our previous work.

Design and operation of an imaging sensor for detecting vegetation

There is a need to sense vegetation from ground‐based vehicles so that plants can be treated in a selective way, thus saving on crop treatment measures. This paper introduces a sensor for detecting vegetation under natural illumination that uses three filters, red, green, and near infra‐red (NIR), with a monochrome charge couple device (CCD) camera. The sensor design and the data handling are based on the physics of illumination, reflection from the vegetation, transmission through the filters, and interception at the CCD. In order to model the spectral characteristics of the daylight in the NIR, we extend an existing standard using a black body model. We derive suitable filters, develop a methodology for balancing the sensitivity of each channel, and collect image data for a range of illumination conditions and two crop types. We present results showing that the sensor behaves as we predict. We also show that clusters form in a measurement space consisting of the red and NIR chromaticities in accordance with their expected position and shape. Presentation in this space gives a good separation of the vegetation and nonvegetation clusters, which will be suitable for physically based classification methods to be developed in future work.

Model-based control of image acquisition

Abstract We propose two methods for controlling the acquisition of images using a camera/digitiser combination which seek to make good use of the dynamic range of the digitiser. The system controls are the black and white reference levels of the digitiser, and the exposure time of the CCD sensor. We use the grey level histogram to characterise the level of control. Both methods use models of the camera/digitiser and of the grey level distribution in the scene. These allow control values that will achieve a given result to be predicted from the current grab and used on the next one. Thus the methods use feed forward control, taking advantage of the models to achieve a fast response. The first method, pragmatic, attempts to adjust the controls to achieve target values of histogram position and scale. The second method, information theoretic seeks to maximise the information content of the histogram as measured by the entropy. An advantage of the information theoretic method is that it produces a single measure of performance. This we use in a strategy for including the exposure variable in the control system. Having a single measure avoids the difficult problem of choosing rather arbitrary weighting factors for the position and scale errors in the pragmatic method. We test both methods using stored images and simulating various grab conditions. Both methods perform well, resulting in effective control values from simulated grabs containing significant saturation. We test the second method on line using real grabs and show fast and accurate recovery from disturbances of illumination and scene content.