• No results found

Unlike natural images which are selected by the photographer to be visually ap- pealing and free from technical errors, microscopic images present often defects due to the automatic acquisition process and the sample preparation. This Chapter prepares the way to further processing of HCS imagery datasets by presenting a framework for microscopic image quality control based on one class learning. In this chapter we study the characteristic of the typical defects which are found in a HCS imagery dataset acquired with a state of the art microscope. Based on our observations, we show that its possible to distinguish global and regional defects with a scalable cascade of one class classifiers and we propose appropriate features for distinguishing these classes. It is worth noticing that our method requires only

2.5. Discussion 27 X Y

12

24

36

48

60

72

84

96

12

24

36

48

60

72

84

96

−0.1 −0.05 0 0.05 0.1 0.15 0.2

Figure 2.10: Location of out-of-focus images on96 × 96 cell array layout. The

color represents the signed distance to the classifier’s decision boundary. Higher value signals a severe out-of-focus error. The prominent strip in the center indicates an error in the focusing system of the microscope during acquisition.

Figure 2.11: Examples of error in the detection of regional defects. Left two, two missed detections of regional defects with a weak intensity. Right two, two false positive detections where dark and slightly out of focus images are detected as regional defects.

training images for the normal class. In this imaging regime the preparation of a representative training set would be too time consuming for fully supervised multi- class learning as it would require mining a large part of the data. Our approach avoid this costly labelling effort.

In the future, multiple extensions of this research are possible. We showed that the running time of our framework is comparable with data acquisition time. This can be a practical benefit as it can allow to correct detected out-of-focus images while the microscope is scanning the well plate. In addition, given the generic structure of the classifier cascade, we would like to extend our framework to other imaging modalities. One realistic scenario is the detection of missing slices, a defect that often occurs in volumetric electron microscopy imaging.

Chapter 3

Learning to Count with

Regression Forest

and Structured Labels

Counting objects in images or video frames is important in many real-world appli- cations including industrial inspection, cytometry, surveying and surveillance. For example, in HCS a common assay to asses the toxicity of a certain drug is to count the number of cells which survive after treatment. If the objects in the image are isolated and their appearance makes them distinct from the background, the simplest approach is to segment the image and count the number of foreground connected components. However, when multiple objects overlap this method provides a poor estimation because several instances can be merged in a single connected component (a cluster). This chapter investigates a recently proposed method for counting ob- jects in static images (density-counting Lempitsky and Zisserman (2010)) that avoids the hard task of segmentation and detection. This approach seeks to count objects by integrating over an object density map that is regressed from local image features.

This chapter introduces a novel density-counting algorithm. Unlike previous work (Lempitsky and Zisserman, 2010)) where the prediction for each pixel is independent, we obtain the density map by averaging over structured, namely patch-wise, predic- tions. Using an ensemble of randomized regression trees that uses dense features as input, we obtain results that are on par with the state of the art methods but require only a fraction of the training time and lower implementation effort. Given the numerous possible applications of this algorithm, an open-source implementation is currently under development in the framework of http://ilastik.org. The software will provide a graphical user interface to faster training of the algorithm on different datasets.

3.1

Related work and contributions

Recent work has shown that object counting can be solved with equal or better accu- racy without prior object detection or even segmentation (Lempitsky and Zisserman, 2010; Ryan et al., 2009; Chan et al., 2008). In fact, such an approach is the only one viable in settings of such crowding or such low resolution so that detection and segmentation of individuals becomes impracticable. In some instances, a count estimate may also boost the performance of object detectors (Rodriguez et al., 2011).

Figure 3.1: Summary of our framework. A regression random forest learns a mapping between patches in the input feature space and in the target object density space. Overlapping predictions of patches are averaged to obtain a density of objects per pixel.

An estimate of the object countNocan either be obtained directly, by mapping from

a set of global features to the real line or the integers (Cho et al., 1999; Kong et al., 2006; Marana et al., 1997); or it can be obtained by integrating an estimated density

functionF(x) over the image domain Ω (density counting)

No=

Z

F(x)dx (3.1)

whereF(x) is computed from local features. The latter approach delivers state of

the art performance while requiring less training images than global regression methods. In particular, the pioneering work (Lempitsky and Zisserman, 2010) posits

Related documents