In one example, the image data is obtained from an existing image which could have, for example, been taken for another purpose. In another example, the image data is obtained from an image taken using an image capture device dedicated to the present method. In any event, the images may be processed to produce the image data, for example a plurality of images including the same microbial growth may be averaged, demosaiced and colour calibrated.
The image data may include a series of such processed images, for example taken using different lighting configurations. Further details of image capture and processing will be given below. The classifier may be a boosted classifier. Alternatively, the classifier may be a decision tree, random forest, or may use Linear Discriminant Analysis LDA or any other technique to classify the pixels. A boosted classifier combines the output of a set of "weak" classifiers with low, but better than random, classification accuracy to produce a single "strong" classifier with high accuracy.
In combining the weak classifiers, the result of each weak classifier is preferably weighted according to the confidence in the correctness of the weak classifier. An example of an appropriate boosting algorithm for creating the boosted classifier is Discrete AdaBoost, which will be described in more detail below. Discrete AdaBoost adaptively improves the performance of the classifier by giving greater weight to examples misclassified by a weak classifier when training the next weak classifier.
The variants use different methods of weighting and training the weak classifiers. In an embodiment, the classifier used to classify the pixels is a boosted decision tree classifier. In this embodiment, the weak classifiers used are decision trees. Decision trees are binary trees where each node contains a rule, each branch represents a result of the rule, and hence a decision, and each leaf representing a classification.
The tree is traversed from the root node to classify a single feature vector. A common arrangement is that if a test passes the left branch of that node is traversed, otherwise the right branch is traversed instead. The boosted decision tree classifier may be a multi-class classifier.
For example, the AdaBoost. MH Multi-Label Hamming procedure may be used for classifying the pixels into one of a plurality of classes. The AdaBoost. For better classification speed and performance, classification may be performed as a two stage process. For example, the method may include using a first classifier to initially classify each pixel as one of a first plurality of classes, and subsequently using a second classifier to classify each pixel in one or more of the first plurality of classes as one of a second plurality of classes.
The result is a chain of dependent classifiers, where the output of the first classifier is used to either augment the feature vectors or restrict the applicable pixels when training the second classifier. The first classifier may be a binary cascade classifier to initially classify each pixel as background or non-background, and the second classifier may be a multi-class boosted decision tree classifier to classify each non-background pixel as one of the second plurality of classes.
The initial coarse classification of background pixels reduces the number of pixels that need to be more accurately classified into one of the colony types.
More work for NMC as local body tax replaces octroi
The initial binary classification may have a high false positive rate background pixels classified as non- background as misclassified pixels can be correctly classified in the next stage. A suitable initial binary classifier may follow the cascade classifier method described in Paul Viola and Michael Jones "Robust real-time object detection", Second International Workshop on Statistical and Computational Theories of Vision, July 13 , the contents of which are herein incorporated by reference.
The classes may represent organism genus, species or strains, agar types, background, non-background.
The classes are flexible, to allow for different agar types and organism types that may be of interest to different end-users. The method may further include assigning a confidence value to each pixel classification. The confidence value represents the probability of correct classification of the pixel. The method may further include applying a post processing algorithm to improve the results of the pixel classifications.
A range of different algorithms may be applied to remove spurious labels or uncertain areas. For example, the post processing algorithm may include morphological operations such as dilation or erosion or, alternatively, a graph cut. Pattern Anal. Graph cut algorithms compute an optimal partitioning of a graph into two or more sets.
The application of the graph cut algorithm may lead to the reclassification of some of the pixels in the image data. Applying a graph cut algorithm may include:. Applying a graph cut algorithm may improve the classification results for pixel classifications that have been given low confidence by the classifier. This spatially smoothes the results, reducing noise in the final classifications. In particular, the high confidence annotations will normally be kept, while low confidence annotations may be replaced to obtain a more homogeneous output.
For example, if a low confidence pixel is surrounded by neighbouring pixels classified into the same class with high confidence, there can be more confidence in the low confidence pixel's classification. By the same token, if the neighbouring pixels are classified into a different class with high confidence, there is a strong case for changing the low confidence pixel's classification to that class.
- The Orchids of Cuc Phuong National Park - lllustrated Guide.
- Grafte op die berg.
The graph cut algorithm may be a multi-class graph cut algorithm, which partitions the graph into three or more sets. This is a more complicated algorithm than a binary graph cut, but allows a more meaningful assessment to be provided as the pixels may be classified more specifically. Such an algorithm may be an alpha expansion procedure.
In alpha expansion, a series of graph cuts are performed, each time segmenting between the current label for each node, and a candidate label from the set of possible labels. This procedure is repeated, iterating through each possible label, until convergence. In constructing the graph, auxiliary nodes are added between adjacent nodes with different labels, to include the cost of this labelling in the cut.
- Manual LBT Bare Rules 2012.
- Social Aspects of Language Acquisition - Language Socialization and Grammatical Development.
- Goodbye, A New Beginning.
Further details of appropriate graph cut algorithms are described in the co-pending International application to the same Applicant and with the same filing date titled "Method for Improving Classification Results of a Classifier", the contents of which are herein incorporated by reference. The feature vector may be composed of values taken from feature images, which are created by applying one or more filters to the image data for the sample. Obtaining feature images is an intermediate step, and it will be appreciated that alternatively the feature vectors could be composed of values taken by applying the filters on a pixel- by-pixel basis.
The term "filter" is to be understood to encompass any image processing or other algorithm that converts part or all of the image data into one or more values that may be input into the classifier. The one or more filters may be chosen from a group including: o RGB filter, which creates up to three feature images, one for the red channel, one for the green channel and one for the blue channel,.
It is also possible to utilise first and second order derivatives of the Gaussian kernel,. Other position filters may show the distance of each pixel from a particular label, if the filter is applied when the output of a previous classifier is available,. The filters may enhance the differences between colonies and the medium and the plate itself. Some of the filters include a feedback mechanism, whereby the output from a previous classifier can be used as an input.
The feature vectors may be constructed using any number of filters. For example, the feature vector for a pixel may include 9 values: Xi , x 2 and x 3 corresponding to the R, G and B channel values for that pixel, x 4 , x 5 and x 6 , corresponding to the L, A and B values for that pixel, x 7 corresponding to the distance of the pixel from the edge of the plate and x 8 and x 9 corresponding to image gradients in the x and y directions. It will be appreciated that any number of feature values may be used and feature values may be taken from different image data for the same growth e.
In relation to texture, applying one or more filters to the image data may include measuring texture of the image data by. The gradient values may be extracted by convolving in x and y directions of the image data using a Sobel kernel. The use of a Sobel kernel smooths the image at the same time as extracting gradient information. Computing the trace of the covariance matrix of the image gradients measures the variability of the gradients in the region and hence the degree of texture, providing a "smoothness" metric.
The computation required during training may be reduced by sparsely sampling the plurality of pixels from the image data. This is especially useful for large images. Two algorithms that may be used include dense or sparse sampling. In training, the sparse sampling method constructs a separate pool for each class being trained, leading to N pixels being available for every class. The dense method constructs a single pool for all of the classes, so the number of pixels in each class will depend on the distribution of that class in the images.
The method may also identify the medium as a negative control.
Reading : Prisons Handbook Edn5
For example, the classifier may classify pixels as background, or non-growth. The method may determine whether the number of pixels classified as background exceeds a predefined minimum number for the image capture and analysis to be valid. The predetermined minimum number of pixels that must be classified as background e. Note that the maximum is simply the number of pixels that lie on the culture plate in the image.
As a further check on the integrity of the image capture and analysis, the confidence of the classifications may be examined. If any regions of sufficient size exist with consistently low confidence, it is likely that an unknown item is on the culture plate. In this case, the operator can be alerted, or the culture plate marked for inspection. To generate the assessment, the results of the pixel classifications are analysed. This may involve processing the pixel classifications to extract additional metadata, such as colony counts, colony sizes, the range of organisms present on the culture plate, growth patterns, haemolysis, artefacts, contaminants, defects in the agar, etc, as described above.
The number of each type of colony is counted enumerated using a process known as quantitation. It will be appreciated by those persons skilled in the art that counting the colonies is a difficult task, since many colonies may not be isolated from each other.
WO2012119191A1 - Method and software for analysing microbial growth - Google Patents
Three approaches are therefore proposed:. Basic counting, where the colony count is estimated by determining an average colony size for each organism type on each type of agar. The number of pixels of each organism detected on a particular culture plate can be divided by this number to yield the number of colonies.
Density based counting, where the distance from the edge of each colony is included in the counting algorithm. This is based on the observation that large areas of growth should be counted as several distinct colonies. Hence interior pixels in the centre of a colony are assigned higher weight to achieve such an outcome.