JP2005032210A - Method for effectively using spatio-temporal image recomposition to improve scene classification - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a method of using many recomposed modifications of an input digital image for the purpose of improving scene classification. <P>SOLUTION: A method for improving scene classification of a digital image comprises the steps of; (a) providing an image; (b) systematically recomposing the image to generate an expanded set of images; and (c) using a classifier and the expanded set of images to determine a scene classification for the image, whereby the expanded set of images provides at least one of an improved classifier and improved classification result. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates generally to the field of digital image processing, and more particularly to a method of using multiple reconstructed variants of an input digital image to improve scene classification.

  It is difficult to automatically determine any image with meaningful classification (eg sunset, picnic, sandy beach, etc.). Much research has been done recently, and various classifiers and feature sets have been proposed. The most common designs in such systems use low level features (eg, color, texture, etc.) and statistical pattern recognition techniques. Such a system is based on a sample that depends on a learning pattern from a training set (see, for example, NPL 1). Such a sample-based system is in contrast to a model-based system in which the type of features are identified directly using human knowledge or a hybrid system in which the model is learned.

  Semantic scene classification can improve the performance of content-based image composition and retrieval (CBIR). Many current CBIR systems allow a user to identify an image and search for images that are similar to the identified image, where similarity is often determined only by color or texture characteristics. This so-called “query by example” has often proved inadequate. Knowing the scene category from the beginning helps to dramatically narrow the search space. For example, knowing the composition of the party scene allows us to consider only the party scene in our search, answering the question “Find Mary's birthday party photo”. Thus, the search time is reduced, the hit rate is higher, and the false alarm rate is expected to be lower.

  Current scene classification systems enjoy the limited success of unconstrained image sets. What seems to be the reason for this seems to be the incredible kind of image found in the most meaningful classification. A sample-based system must account for such changes in the training set. Hundreds of samples do not necessarily capture all the variability inherent in some types. As a specific example, when sunset images are classified, sunset images taken at various stages of sunset tend to glow more brightly as the sun approaches the horizon, and then tend to fade over time. Because there is a color can be very different. Furthermore, does the configuration include only the horizon or the sky? Where is the sun in relation to the horizon? Is the sun in the center or offset to one side? Is not more uniform in the field of view of the camera.

  A second reason for limited success in sample-based classification is that the image appears to be less prototypical to the scene, and therefore does not match well with any of the training samples, excessive foreground regions or confusion It often includes the foreground area. For example, FIG. 1 shows four scenes (a) through (d) with a foreground area to be confused. This is particularly true in camera user images when a typical camera user pays less attention to composition and light than a professional photographer. Thus, the camera user's images cause a great deal of variability when used in this area, causing high performance in many existing systems (with professionally photographed stock photo libraries such as the Corel database) including.

  Consequently, there is a need to provide a method that overcomes the problems described above in image classification. These problems were proposed by the introduction of the concept of spatial image reconstruction, designed to minimize the impact of improper configurations (ie, foreground objects), and are also simulated or effective. Designed to minimize the impact of temporary image reconstruction and to minimize the color change effects that occur over time.

This approach is supported by past successes in other areas. In face recognition and detection, researchers used confused deformations of faces in training to handle geometric changes (see, for example, Non-Patent Document 2). This is related to resampling or bootstrapping. In addition, bagging (aggressive bootstrapping) uses multiple variants of the training set to train different component classifiers, and the final classification decision determines the classifiers in the individual components. Based on the selection (for example, see Non-Patent Document 3).
A. Vailaya, M.M. Figueiredo, A.A. Jain, H.J. J. et al. Zhang, "Content-based hierarchical classification of vacation images", Proceedings of IEEE International Conference on Multimedia Computing and Systems, 1999. H. Rowley, S.M. Baluja, T.A. Kanade, "Rotation invariant neural network-based face detection", Proceedings of IEEE International Conferencing on Computer Vision and Pattern Recognition, 1998. R. O. Duda, P.M. E. Hart, D.D. G. By Stock, “Pattern Classification”, John Wiley & Sons, New York, 2001, pp. 475-476 R. P. W. Duin, "The combining classifier: To train or not to train?", Proceedings of International Conference on Pattern Recognition, 2002. B. Scholkopf, C.I. Burges, and A.A. Smol, "Advanceds in Kernel Methods: Support Vector Learning", MIT Press, Cambridge, MA, 1999, pp. 263-266 Y. Wang and H.W. Zhang, "Content-based image orientation detection with support vector machines," Proceedings of IEEE Workshop on Content-Based Access of Images 200.

  It is an object of the present invention to provide a method that uses multiple reconstructed variants of an input digital image to improve scene classification, leading to overcoming one or more of the problems described above. .

  According to one aspect of the invention, (a) providing an image, (b) systematic reconstruction of images to generate an expanded set of images, and (c) image expansion. Using a classifier and an expanded set of images to determine image classification in the image, wherein the set provides at least one improved classifier and improved classification results Provide a way to improve classification.

  The present invention provides a method for systematically generating a reconstructed variant of a sample image to generate an expanded set of training samples that derive a robust classifier, or the same salient properties that derive a robust image classification result. Either (or both) methods are provided for systematically generating a reconstructed variant of an input digital image to be tested to generate an expanded set of test images. This has the advantage of providing a way to increase the diversity of training samples, accept a better match between samples and images, and obtain a more robust image classification.

  The advantages of the present invention can provide a method that uses multiple reconstructed variants of the input digital image to improve scene classification.

  The invention will be described as executing on a programmed digital computer. Those skilled in the art of digital image processing and software programming will be able to program a computer to carry out the invention from the description below. The invention may be embodied in a computer program product having a computer-readable storage medium, such as a magnetic or optical storage medium that holds machine-readable computer code. Alternatively, the present invention may be implemented in hardware or firmware.

  A major obstacle to high performance in meaningful scene classification is the vast change in both the color and composition of images in each type. Obtaining sufficient training data for a sample-based system may be a difficult task, especially if the type contains many changes. Manually collecting a large number of high-quality and original images is time consuming even with the help of a theme-saved photo library. Therefore, it is important to use all available training data efficiently.

Furthermore, if the image matches that type of sample in both its color and its configuration, the best match of the test image with the set of training samples occurs. However, the test image may include changes that are not present in the training set. The degree of matching is affected by the image picked by the photographer (which affects the image composition) and the time taken (potentially affecting the color of the image due to changes in scene illumination over time). If it is possible to “reproduce” the scene, it would be possible to attempt to obtain an image with a more original color and composition in kind. For example, referring to FIGS. 2 (a) to 2 (d), the original scene (FIG. 2 (a)) contains the salient sub-regions that were cropped and resized (FIG. 2 (c)). Finally, in FIG. 2 (d), the illumination shift is applied simulating the subsequent sunset. How can you "reproduce" a scene? In other words, how can any image be converted to an image that will match the original sample well?
According to the present invention, a concept called effective space and temporal reconstruction is used to present the above problem. In general, image reconstruction is defined as a process that systematically creates modified variations of the same image, including spatial reconstruction and color composition. The different types and uses of spatial reconstruction (reflection and cropped images) and effective (simulated) temporary reconstruction (image color shift) are represented in Table 1 and are described in more detail below. Explained. They are categorized as training, testing, and reconstruction in both. Numerous types and combinations of use require visual inspection to ensure that such reconstruction does not destroy the integrity of the training sample (eg, aggressive cropping is a loss of the main subject of the picture You may end up in.)

トレーニングでの再構成
トレーニングデータの限定されたサイズセットで再構成を用いることは、見本のより豊富で、より多様なセットを生じうる。目標は、各画像を視覚的に検査する必要なしに、これらの見本を得ることである。１つの技術は垂直軸に関する各画像を反射することで、それによって、見本の数を倍にする。例えば、図３（ａ）乃至３（ｃ）に示されるように、オリジナル画像（３（ｂ））は、水平な反映（３（ａ））又はクロップ（３（ｃ）に示されるように底部から２０％）によって変換される。明らかに、新規画像の分類は変っておらず、すなわち、画像の左側の太陽の夕焼け画像が右側に太陽を移動させる一方、画像は有効な夕焼け画像のままである。

Reconstruction with training Using reconstruction with a limited size set of training data can result in a richer and more diverse set of samples. The goal is to obtain these samples without having to visually inspect each image. One technique is to reflect each image about the vertical axis, thereby doubling the number of samples. For example, as shown in FIGS. 3 (a) to 3 (c), the original image (3 (b)) can be viewed in the horizontal reflection (3 (a)) or crop (3 (c) as shown in FIG. To 20%). Clearly, the classification of the new image has not changed, i.e. the sunset image of the sun on the left side of the image moves the sun to the right, while the image remains a valid sunset image.

  Another technique is to crop the edges of the image. It is assumed that the prominent part of the image is at the center and the incomplete composition is caused by the surrounding mess. Cropping sequentially from each side of the image results in four new images of the same classification. Naturally, you don't want to lose a noticeable part of the image, but for example, in a small conservative crop of 10%, the algorithmic classification may change, but the meaningful classification of the scene changes. Is unlikely.

Reconstruction in the test Reconstruction of the training set gives a further example, but reconstructs the test image and classifies each new, and the reconstructed image yields a composite classification of the original image. With respect to spatial reconstruction, the edges of the image can be cropped for the purpose of better matching the test image features to the sample. To obtain such a match, it may be necessary to crop more aggressively (as shown in FIG. 2). However, if the classifier trains with the reflected image, it is not necessary to reflect the test image due to the symmetry already built into the classifier. For example, if a 1-NN classifier is used, the test image feature vector T would be located at a distance from the nearest example vector E. Call the vectors of the reflected images E and T, E ′ and T ′, respectively. Due to the symmetry of the feature, d (E, T) = d (E ′, T ′), and T ′ is redundant.

  Some types of images contain a great deal of change in color distribution in the image type world, and changing the overall color of the test image may properly produce a good match with the training sample. Absent. Using the sunset image classification as an example, the early sunset and the later sunset may have the same spatial distribution of colors (bright sky over dark foreground), but early The overall look of is more chilly due to the color change in the lighting of the scene. By artificially changing the color along the luminance (red-blue) axis towards the warmer side, we can simulate the look of the image that will be captured later, and we will shift this to illuminate This is referred to as effective temporary reconstruction. For example, as shown in FIGS. 4 (a) to 4 (f), if the button is equal to a photo stop of 0.4, the temporary reconstruction starts with the -6 button (FIG. 4 (a)). With a series of illumination shifts in 3 button increments, ending with the +9 button (FIG. 4 (f)). Similarly, changes in light intensity in a scene can be handled using changes along the luminance axis. Color changes along other axes may be applied in other problem areas.

  Using either spatial or temporal reconstruction, the classifier may or may not label the new reconstructed image with the same type as the original image. How do you determine when the reconstructed images have different classifications? Duin discusses two types of combiners, fixed and trained (RPw Duin, “The combining classifier: To train or not to train?”, Proceedings of International Conf. on pattern recognition, 2002). Fixed combination rules include scheme selection and use of sum or average of scores. A trained combiner is a second classifier for mapping scores to a single score. Two considerations affect the availability of training data and the choice to use the degree to which the underlying classifier is trained. Duin proposes that the classifier being trained can benefit from the trained combiner while not being able to overtrain (eg, support vector machines (SVMs)). In today's studies, this has been found to be the case (eg, second stage SVM does not support).

  In two types of problems, the fixed combiner of interest for r reconstruction is the m th order statistic, eg maximum (m = 1), second largest (m = 2), or intermediate (m = r / 2) is used. A change in parameter m moves the position of the classifier in the operating curve. A small m gives a larger recall at the expense of more error positives and actively classifies images in a proactive way. The choice of m will obviously depend on the application.

  Furthermore, the scores can be combined in a way that finds the most consistent image classification. For example, scheme boats can be used for combination. This is desirable, and classification based on a number of slightly modified reconstructed images with the same salient scene content should be more robust than classification based only on the original image. A single classification based on the original image is inaccurate due to some statistical irregularities (eg, insufficient foreground confusion with foreground confusion or sample set), and many more reconstructed images Classified correctly, the majority of rules will correct irregularities.

Reconstruction in both training and testing In many applications, reconstruction may be used in both training and test data. They can be easily combined so that each serves a different purpose. You may ask if you need to use both types of reconstruction. That is, why do we need to reconstruct test images if we have a sufficiently rich set of training samples? The need to use reconstruction in both training and testing is practical. The training data is sufficiently diverse to start, or there is no guarantee that the reconstruction of the training sample has completely created any changes and completely fills the image space.

  A related issue is the choice between training image reconstruction and acquisition of additional unique samples. Apart from the initial discussion on the lack of good training data and the time required to collect the data, there are more effective data property issues. Reconstructing a small set of prototype samples would be more desirable than a more used but poor quality sample.

  In addition, the use of reconstruction in the test at the top of the reconstruction in training is a way to certainly boost recall if desired, at the expense of potentially high false alarm rates.

  The last question is whether a more aggressive approach is used to reconstruct the training data to minimize the need to reconstruct the test data. Since aggressive reconstruction can lose significant content from the image, it should be ensured that the integrity of the expanded training set is not compressed, and the discussion is a technique for doing this accurately.

Semi-managed reconfiguration in training Our purpose of using conservative reconfiguration in a training set is to do a process that is not fully managed. However, if more training data is desired and aggressive reconstructions such as large crops or significant color shifts are used, there is no need to go to the opposite pole to inspect all the reconstructed images. In addition, a training methodology is required.

  Obviously, since some aggressive reconstructions can move certain scene content characteristics to image types, the exact approach to adding these images to the training data is the reconstructed training image Each would be visually inspected. Doing so can be tedious and cumbersome. It would be more efficient to only examine a subset of the reconstructed image that requires attention. To reflect the reconstructed image, the classifier can be trained using the original training image, and the classifier is then used to classify the reconstructed form of the training image. Only the reconstructed images that fail the classifier (or pass with low confidence) need to be visually evaluated in order to determine when the reconstruction causes significant scene content to be lost from the images. Such reconstructed images are then deleted, while the remaining reconstructed images are added to the expanded training set to improve its richness. This is a favorable exchange between the generation of a small number of reconstructed images in an unmanaged manner and the generation of a large number of reconstructed images in a fully managed manner.

  Next, the three embodiments of the present invention, preferred sunset detection, outdoor scene classification, and automatic image configuration detection are each described.

Sunset Detection With the hierarchical image classification scheme described by US Pat. Confirming the intuition that sunsets are recognizable by their glowing warm colors, the edge direction was felt more prominently in such problems than other features, as the edge direction alone. In addition, spatial information should be incorporated to distinguish sunsets from other landscapes that contain warm colors such as desert rock formations. Thus, spatial color moments may be used by dividing the image into 49 regions using a 7x7 grid and calculating the average and change of each band of the Luv transformed image. This yields 49x2x3 = 294 features.

Support vector machines (SVM) are similar problems (for example, B. Schokopf, C. Burges, and A. Smola, Advanceds in Kernel Methods: Support Vector Learning , MIT Press, 19Mandp. , And Y. Wang and H. Zhang, “Content-based image orientation detection with support vector machines,” Proceedings of IEE Workshop on Content-Based Accord. Tyzor to (Learning Vector Quantizers) (LVQ) for (Learning Vector Quantizers) outperforms other classifiers such as are preferably used as classifier. In particular, a Gaussian kernel that generates RBF style classifiers (see RBF = Radial Basis Function, Wang and Zhang) was used. SVM is designed for two classes of problems and outputs the actual number in each test image. Signs are classifications, and size can be used as a means of loose confidence.

  Since the set was much richer, using resynthesis in the training set significantly increased the performance significantly. This overcomes a number of effects with a limited training set. Using reconstruction in the test set increased both the number of strokes and the number of false positives. Finally, using reconstruction in both training and testing gave the best overall results. Note that the results correspond to the optimal operating points on the different curves.

  Using spatial reconstruction in the test image achieved the goal of accurately classifying sunset images with foreground regions that were rather confusing. For example, the images displayed in FIG. 1 were all incorrectly classified by the baseline system, but were correctly classified when reconstruction was used (obtained by reconstruction). The upper right image (b) is a good example of how reconstruction by cropping can help. Cropping the vast and dark water areas of the foreground in the image substantially increases the SVM score. Do the same for the other images, for example, clipping the bottom 20% from the bottom of the left image (a) will remove the confusing reflections in the water.

  However, the number of false positive images also increased, partially offsetting the increase in recall. A typical error positive caused by reconstruction is shown in FIGS. 5a and 5b. Each of these images includes an atypical pattern of sunset (eg, complex bright areas in the night view or sky in the desert landscape) that, when cropped, produces an image that more closely resembles the sunset.

  Many sunset images have an original configuration, but have a weak color corresponding to an early or later sunset. Shifting these images "warming up" in the scene illusions causes them to be classified correctly, and they both introduce many false positives as shown in FIG.

Classification of outdoor landscapes The system described above is extended to identify six types of outdoor landscapes: sandy beaches, sunsets, fallen leaves, fields, mountains, and cities (defined in Figure 2). Images used for training and testing included images of Corel and camera users. The SVM classifier is a one-to-all approach (see B. Scholkopf, C. Burges, and A. Smola, Advanceds in Kernel Methods: Support Vector Learning , MIT Press, 99, MIT Bridge, 99. The same features and classifiers are used as for sunset detectors. Spatial reconstruction was particularly effective when used in training, as the training set is still limited. Reconfiguration was not used in the test set.

Image Coordination Detection Automatic Image Coordination Detection (Y. Wang and H. Zhang, “Content-based image orientation with detection of the first and third of the worlds,” Proceedings of IE The purpose of (see) is to classify any image into one of four magnetic needle directions (N, S, E, W), depending on the direction the top of the image is facing. Doing so based solely on image content is a difficult problem. In a preferred embodiment, a baseline system that achieves similar results similar to Wang et al. Uses spatial color moments and one-to-all SVM classifiers.

  Reconstruction in the test can be expected to improve the classification of this region as well, but the logical basis for using that prediction is very different: the cropping of the image edge is the perceived configuration of the image Should not affect. Therefore, combination classification based on a number of slightly different images should be more robust than single image classification. Testing with both fixed (choice) and trained combiners, it was found that the performance of each was comparable, and the choice was chosen for performance simplification.

  In this application, images are classified with 4 scores each from SVMs adjusted to recognize images of a given configuration. One-to-all classifiers classify images with a configuration corresponding to the SVM that produces the highest score. This process is repeated nine times, once for each cropped deformation of the image. Such a process uses the score to break the bond and finally selects in 9 classifications (bonding does not mean a single coordinated domination, but the image is a good candidate for rejection, ie No obvious coordination). An example of a selection scheme is given in FIG.

  A sample collel image obtained using the reconstruction scheme is shown in FIG. In each of them, the many regions at the image boundaries are confused. Dark shadows (Fig. 8 (c)), dark trees (Fig. 8 (b)), and reflections from the sun (Fig. 8 (c)) all confused the classifier, with bright or dark areas in the image. Appears on the side, not up or down.

  Image reconstruction is similar to the spirit of the bootstrap or bag method, with the main distinction that only a single classifier is trained and used in classification. The key to the successful application of this scheme to the image classification problem is that such image reconstruction can ignore those classifications in the final classification, and significant content is invariant to such confusion to the image It will simply affect the components that are confused in the image in such a way. This is therefore a general approach to boosting classification performance as long as the appropriate method of reconstruction is selected by the problem area and the feature / classifier used.

  The following guidelines are presented to assist in determining how to use image reconstruction in image classification. First, if the training set is sparse, the use of conservative and spatial reconstruction may be very helpful. More aggressive reconstruction of both spatial and temporal should be done in a semi-controlled manner. In the two types of problems, the reconstruction of the test image gives operational curve parameters that can be used to modify the performance to the application, causing a good match with the same type of sample. For many types of problems, selecting during classification of the reconstructed image is more robust. Obviously, in the ideal case where the types are well separated in the training data and the test images are in good agreement with the sample, the reconstruction is not expected to greatly assist.

  FIG. 9 illustrates a method for improving scene classification of digital images according to the present invention. First, either the sample image 10 or the input test image 12 is provided to the input stage 14 and then, as described in the detailed description of the invention, a spatial reconstruction algorithm 18 or a temporal reconstruction algorithm 20. Applied to the reconstruction stage 16, in which the image is systematically reconstructed according to either (or both). The result of the reconstruction depends on the type of input image (sample or test image), which may be either an expanded set of sample images 24 or an expanded set of test images 26 (or both). Is an extended set of If the expanded set of images is a sample image, the expanded set of images is used to train with the classifier at training stage 28, thereby providing an improved classifier according to the present invention. If the expanded set of images is a test image, the expanded set of images is used in the classifier stage 30, thereby providing improved image classification results according to the present invention. The improved classifier resulting from the expanded set of sample images 24, as indicated by the dotted line 32 connecting the training stage 28 and the classification stage 30, is to provide improved overall classification results. , May be used with an expanded set of test images 26. In addition, however, it is possible to apply the reconstruction stage 16 to only one of the two paths shown in FIG. 9 (i.e., improved classifier training or providing improved image classification results). Or both.)

  The subject of the present invention is to recognize and select useful meanings for human-understood objects, properties or conditions and then digitally convert the digital image to take advantage of the results obtained in further processing of the digital image. Related to digital image comprehension technology, which is understood to mean a technical processing technique.

  Scene classification can also be seen in applications in image enhancement. Rather than applying an overall color balance and exposure adjustment to every scene, the adjustment may be customizable for the scene. For example, while removing a warm cast from an indoor image illuminated with tungsten, the glowing color in the sunset image is retained or boosted.

  The reconstruction technique described by the present invention is not limited to photographic images. For example, spatial reconstruction can also be applied to medical images in medical image classification (although color reconstruction does not apply).

One of the four images of a sunset image with a confusing foreground area is shown. One of the four images of a sunset image with a confusing foreground area is shown. One of the four images of a sunset image with a confusing foreground area is shown. One of the four images of a sunset image with a confusing foreground area is shown. FIG. 6 shows one of a series of images that illustrate how an arbitrary image (a) is transformed into an image (d) that is in good agreement with a prototype sample. FIG. 6 shows one of a series of images that illustrate how an arbitrary image (a) is transformed into an image (d) that is in good agreement with a prototype sample. FIG. 6 shows one of a series of images that illustrate how an arbitrary image (a) is transformed into an image (d) that is in good agreement with a prototype sample. FIG. 6 shows one of a series of images that illustrate how an arbitrary image (a) is transformed into an image (d) that is in good agreement with a prototype sample. FIG. 6 shows an example of spatial reconstruction when the original image (b) is deformed by horizontal reflection (a) or crop (20% from the bottom as shown in (c)). FIG. 6 shows an example of spatial reconstruction when the original image (b) is deformed by horizontal reflection (a) or crop (20% from the bottom as shown in (c)). FIG. 6 shows an example of spatial reconstruction when the original image (b) is deformed by horizontal reflection (a) or crop (20% from the bottom as shown in (c)). FIG. 5 shows an example of a temporary reconstruction consisting of a series of illumination shifts. FIG. 5 shows an example of a temporary reconstruction consisting of a series of illumination shifts. FIG. 5 shows an example of a temporary reconstruction consisting of a series of illumination shifts. FIG. 5 shows an example of a temporary reconstruction consisting of a series of illumination shifts. FIG. 5 shows an example of a temporary reconstruction consisting of a series of illumination shifts. FIG. 5 shows an example of a temporary reconstruction consisting of a series of illumination shifts. FIG. 6 illustrates an exemplary embodiment of the positive of errors caused by using spatial reconstruction. FIG. 6 illustrates an exemplary embodiment of the positive of errors caused by using spatial reconstruction. How to handle sunset and error positives when temporary reconstruction is used when the original (left image) and shift (+6 buttons) image (right image) is shown FIG. How to handle sunset and error positives when temporary reconstruction is used when the original (left image) and shift (+6 buttons) image (right image) is shown FIG. How to handle sunset and error positives when temporary reconstruction is used when the original (left image) and shift (+6 buttons) image (right image) is shown FIG. How to handle sunset and error positives when temporary reconstruction is used when the original (left image) and shift (+6 buttons) image (right image) is shown FIG. How to handle sunset and error positives when temporary reconstruction is used when the original (left image) and shift (+6 buttons) image (right image) is shown FIG. 2 is an image of one of intentionally confused images showing a winter scene where the sun is close to the horizon but does not sink. How to handle sunset and error positives when temporary reconstruction is used when the original (left image) and shift (+6 buttons) image (right image) is shown FIG. 2 is an image of one of intentionally confused images showing a winter scene where the sun is close to the horizon but does not sink. A table that resolves independent reconstruction decisions using a selection, eg, “T10” means 10% clip from the top of the image. FIG. 4 shows an example of a sample test image obtained by using the reconstruction according to the invention. FIG. 4 shows an example of a sample test image obtained by using the reconstruction according to the invention. FIG. 4 shows an example of a sample test image obtained by using the reconstruction according to the invention. FIG. 2 schematically illustrates elements of a method for carrying out the present invention.

Explanation of symbols

10 Sample Image Input 12 Test Image Input 14 Input Phase 16 Reconstruction Phase 18 Spatial Reconstruction Algorithm 20 Temporary Reconstruction Algorithm 22 Extended Set of Images 24 Extended Set of Sample Images 26 Tests Extended set of images 28 Training stage 30 Classification stage 32 Dotted line

Claims (3)

A method for improving image classification of digital images,
(A) an image provision stage;
(B) a systematic reconstruction step of the image to generate an expanded set of images; and (c) at least one improved classifier and improved classification of the expanded set of images. Using a classifier and an expanded set of images to determine an image classification in the image that provides a result.   The method of claim 1, wherein step (b) includes spatial reconstruction of the images to produce an expanded set of spatially reconstructed images.   Step (b) includes a temporal reconstruction of the images to generate an extended set of temporarily reconstructed images, whereby the images of the expanded set are initially The method of claim 1, further comprising simulating the appearance of late imaging.

Images