CN106203496B - Hydrographic curve extracting method based on machine learning - Google Patents

Abstract

The invention discloses a hydrological curve extraction method based on machine learning. When the method of the present invention performs curve extraction on the hydrological data image, some features with discriminative ability in the image are selected and extracted, and a sampling window with variable scale is used to sample the image pixels in a certain area, as sample data, through machine learning. The method divides image components with different characteristics, and adds new training samples incrementally according to the classification effect; and uses chain code tracking for post-processing, which effectively removes the noise after classification. Compared with the prior art, the present invention solves the problem that the target curve is broken when the hydrological curve to be extracted is thin, which is difficult to be effectively solved in the original hydrological curve extraction method.

Description

Extraction method of hydrological curve based on machine learning

technical field

The invention relates to an image extraction method, in particular to an extraction method of a hydrological curve in a hydrological data image, and belongs to the field of image segmentation.

Background technique

In today's information and digital age, with the popularization of computers and the rapid development of storage media, various research fields are paying more and more attention to the digitization of data and information. For historical reasons, most fields such as hydrology and water conservancy use grid drawings to record observation data. However, due to improper storage and other reasons, paper materials will cause problems such as damage and pollution, and it is easy to cause losses to the information they carry. In addition, paper materials occupy space, and it is not easy to exchange and transmit information, and it is more likely to bury the knowledge that may be hidden and yet to be discovered in the massive information. Therefore, it is necessary to digitize these paper materials. Using image processing to collect these information and establish a database will avoid a lot of manual repetitive labor, and can also enter this information efficiently and accurately, which has strong practical application value.

Paper-based hydrological data are usually blue-purple hydrological curves drawn on orange-red coordinate grid paper. In the process of digitization, each intersection of the hydrological curve and grid lines needs to be obtained when acquiring the information in the drawings, as each moment. observed value. This process requires image segmentation, involving grid line segmentation and hydrological curve segmentation.

Image segmentation is the technology and process of dividing an image into several specific regions with unique properties according to certain standards and extracting objects of interest from them. Image segmentation is the key premise of image analysis, and the quality of its segmentation largely determines the effect of subsequent image analysis. Image segmentation can be divided into grayscale image segmentation and color image segmentation. Compared with grayscale images, color images contain not only brightness information, but also various color information, and their segmentation methods are more diverse, but the corresponding segmentation difficulty is also greater. So far, researchers at home and abroad have carried out a lot of research in the field of color image segmentation, and have proposed many segmentation algorithms and segmentation strategies for specific images, including histogram-based threshold method, region-based method, and edge detection method. , fuzzy clustering segmentation method and neural network method.

In previous studies, the segmentation of hydrological data images is usually based on the threshold analysis method based on color histogram, and the fusion of gradient information and color information is also considered. This kind of method can adaptively complete image segmentation in general, and can reduce the influence of uneven illumination of camera shooting. However, in the actual use of this kind of method, it is found that the hydrological curve obtained by extraction is prone to breakage in some special cases, and the breakage is often very serious, which is difficult to solve by the expansion method.

SUMMARY OF THE INVENTION

The purpose of the invention is to provide a hydrological curve extraction method for the digitization of paper hydrological data, which can accurately extract the hydrological curve and effectively avoid the problem of curve disconnection.

In the hydrological curve extraction method of the present invention, the involved hydrological data images are obtained by photographing paper hydrological data.

The present invention specifically adopts the following technical solutions to solve the above problems.

A method for extracting hydrological curves based on machine learning, comprising the following steps:

Step A: Select the scale of the sampling window and the target feature to be sampled, and collect a representative training sample set accordingly; the scale of the window is scalable. The choice of window scale determines the amount of data used for classification, and also directly affects the scale of computation.

Step B, using the method of machine learning to train and generate a classification prediction model from the training samples;

In step C, each pixel in the image to be processed is collected according to the sampling window to obtain the target feature as the sample to be classified, and the classification prediction model obtained by the training in step B is used for classification;

Step D: Determine whether the classification result of each pixel in the image to be processed is good, so that the curve extraction is complete and there are no other regions with obvious classification errors. If so, enter step F; otherwise, enter step E;

Step E. Select representative sample points from the broken line area of the curve and the area with obvious classification errors, add them to the training sample set after sampling, and repeat steps B-D;

In step F, post-processing is performed on the processed image to remove possible noise.

Preferably, the training sample set collected in step A should at least include samples of three categories: "hydrological curve", "grid line" and "other background".

Preferably, the image post-processing method used in step F adopts a combination of chain code tracking and dilation processing. Before the chain code traces the hydrological curve, the grid line is traced first to determine the mapping area corresponding to the grid line; this step can reduce the processing intensity when tracing the hydrological curve.

Preferably, after chain code tracking, the connected domain whose size is smaller than a certain threshold is regarded as noise, and the image is eliminated. The threshold value is 10000.

Preferably, the size of the connected domain is represented by the area of the smallest circumscribed rectangle of the connected domain.

Compared with the prior art, the present invention has the following beneficial effects:

1. The present invention can better solve the problem of wire breakage that is easily generated when the thin wire is extracted;

2. The present invention is based on the classification of sample patterns, as long as sufficient training samples are selected, the influence of lighting and other issues does not need to be considered;

3. Using offline learning, it is not necessary to re-collect samples for training on each image.

Description of drawings

Figure 1, Figure 2 and Figure 3 are three hydrological data images obtained by shooting.

Fig. 4a and Fig. 4b are the results of extracting the hydrological curve of Fig. 1 and Fig. 2 by the existing method.

Fig. 5a and Fig. 5b are the results of classification prediction of Fig. 2 in different stages of training the classification model in the method of the present invention.

Fig. 6a and Fig. 6b are the classification prediction results of Fig. 3 in different stages of training the classification model in the method of the present invention.

Fig. 7a and Fig. 7b are the results of extracting the hydrological curve of Fig. 2 and Fig. 3 by the method of the present invention.

Figure 8 is a flow chart of the present invention.

Detailed ways

The technical scheme of the present invention is described in detail below in conjunction with the accompanying drawings: Fig. 1 and Fig. 2 show two hydrological data images respectively, and the key to digitizing them lies in the hydrological curve (blue-purple) and the coordinate grid line (orange) among them. red) extraction. It can be seen from the picture that due to the long storage time and the unsatisfactory storage conditions, in addition to wear and tear and paper aging, there are also problems such as color blooming and fading in the image, and even the color shades of the unmapped areas on the same drawing different. And due to the influence of lighting during shooting, the color information in some areas is weakened or loses its original features, which makes the extraction problem more complicated.

The threshold analysis method based on the color histogram adopted in the previous study, which integrates gradient information and color information, can solve the above problems well. The processing results of Figure 1 and Figure 2 are shown in Figure 4a and Figure 4b. It can be seen that the extraction result of the image is good, and the identification and extraction of the target can be roughly completed, but sometimes the extracted hydrological curve will have a broken line, which is a special case, such as the broken line in Figure 4b. After careful observation of these broken lines, it is found that the main reason for the broken lines is that the hydrological curve in the figure is too thin, which makes the ink color of the curve when there is a lot of overlap with the grid line (usually when the curve is approximately parallel to the grid line). Colors that don't completely cover the grid lines will end up showing a superposition of the two. The superposition of this color leads to the migration of color information, where the pixels no longer satisfy the common color characteristics of the hydrological curve, and it is easy to break the line when extracting the curve. What's more, when a certain segment of the curve is approximately parallel to the grid line, there will be a large number of overlapping areas, where the line is often broken, and it is difficult to complete it by methods such as expansion. Since the threshold feature is no longer satisfied, the original processing method is no longer applicable, and adjusting the threshold will lead to a large increase in noise points and unstable image extraction. So consider adopting a new idea to extract the hydrological curve.

Considering that the result of the migration of the color information in the overlapping area is not the same as the existing information such as curve features or grid line features, the goal is to still identify these changed feature information as the target curve. The invention proposes a method for extracting hydrological curves based on machine learning. By sampling pixels in an image for multi-feature fusion, a certain number of labeled training samples are obtained, and a classification prediction model is obtained by training with the machine learning method. Using the model, the pixels in the picture can be classified and predicted, and the pixels predicted as hydrological curves can be extracted. Moreover, by resampling the misclassified pixel area, the training sample set can be made more complete, and the training produces a more robust classification prediction model. In order not to cause the noise generated in the process to affect the curve extraction results, the method of chain code tracking is also used to post-process the extracted hydrological curve images.

Specifically, the present invention comprises the following steps:

Step A, determine the combination of window scale and sampling feature, and collect training samples;

When collecting training samples, it is necessary to determine the scale of the sampling window and the features to be sampled. The sampling window takes the current pixel as the center point of the window, and also takes into account the information of other pixels in the window surrounding the current pixel, that is, local information. Since the information stored in a single pixel is quite limited, when other local information around the pixel is taken into consideration, the sample dimension can be made higher, which is more conducive to detailed classification. The window can use a variety of scales, such as 3*3, 5*5, 7*7, etc. The larger the scale, the more information contained in the sample, which is helpful for more detailed and accurate classification, but the calculation time is also longer; The smaller the sample, the less information in the sample, and the correspondingly, the shorter the calculation time. The specific scale selection should be determined by the needs of the actual application.

On the other hand, the local information used for window sampling also needs to be selected, which consists of some feature values, including color features (RGB, Lab or HSI, etc.), gradient features, texture features (LBP or Gabor, etc.), SIFT features, etc. The selection of specific feature combinations should consider the influence of these features on the extraction effect of hydrological curves, and select appropriate groups of features for combination. Too much feature selection will bring information redundancy and computational load, while too few features may reduce the classification effect. The choice of feature combination is related to the classification effect and the calculation load.

When the above sampling is performed on a pixel, each feature value should be acquired in a certain order according to the pre-agreed window size and feature combination method and integrated into an ordered sample vector. When collecting color features, each pixel in the window should be collected sequentially from left to right and top to bottom. In addition, an additional one-dimensional feature is added to the training sample as the category label of the current training sample.

When collecting training samples, the selection of sampling points is particularly important. It should be noted that: 1. Different target categories should be taken into account when sampling, and each should obtain a sufficient number of sample points; 2. Among the pixels with the same target category, Pixels with different local features should be covered as much as possible; 3. Several typical pixels should be selected for sampling among the pixels of the same category and similar local features. Among them, the collected training sample set should contain at least three categories of samples of "hydrological curve", "grid line" and "other background".

Step B, using machine learning method training to generate a classification prediction model;

Machine learning methods include supervised learning and unsupervised learning. Since the classification goal in the current problem is clear, and only the hydrological curve is expected to be extracted, a supervised learning method is adopted, and the classification prediction model is obtained by using the collected training samples with class labels. Such learning methods include decision trees, Bayesian classifiers, K-nearest neighbors, BP neural networks, perceptrons, and support vector machines (SVMs). Different machine learning methods have different characteristics and should be selected according to actual needs. The choice of machine learning method is related to the effect and efficiency of curve extraction.

Analyzing different images of hydrological data, it is found that the color and structural features of each image are very similar. From the perspective of feature space, even different images can be roughly classified and extracted by the same component interface on the feature space. Therefore, it was decided to adopt the offline learning method. The goal is to train a classification prediction model with excellent effect, which is used to classify and extract the pixels in all the images to be processed, instead of generating a model for each image training.

Step C, classify and predict the image to be processed, and supplement the training sample set;

According to the method agreed in the above-mentioned step A, the corresponding feature samples are extracted pixel by pixel from the image to be processed, and used as input to perform prediction classification through the obtained classification prediction model. The pixels predicted to be "hydrological curves" are extracted as the result of this curve extraction.

If the curve extraction result is complete and the effect is satisfactory, you can proceed to the next step; but usually a satisfactory extraction result cannot be obtained immediately, and the extracted curve is often rough and has broken lines, and many noise points are also extracted. The solution is to continuously acquire new samples to add to the training sample set in an incremental manner, so as to train a more complete and robust classification prediction model. Each new training sample as an increment comes from the previous prediction and classification result, that is, find out the broken line of the curve and other regions with a large classification error rate, and select the typical pixels in the region for sampling. . This method aims to learn to make up for past wrong predictions and resample near the category interface, thereby compensating for omissions and vacancies in the sample space, and obtaining a more complete training set, thereby obtaining a more precise and accurate interface and more Predictive models for robust classification.

Use the training sample set after adding increment to retrain the classification prediction model and perform classification prediction on the image again. If the curve extraction effect is satisfactory, the image passes the current processing and goes to the next step; otherwise, repeat the above incremental training sample process.

The training process of the classification prediction model is not achieved overnight, and requires multiple modifications to add new samples for retraining; at the same time, it is not completed in a certain “training stage”, but only when the classification effect of an image is not good. Initiate "retraining"; that is, without an explicit and limited "training phase". In addition, the addition of the training sample set requires manual intervention, and the newly added sampling points are manually selected. However, since the initial accumulation stage of training samples can usually be completed quickly, a model with better effect can be obtained, and the existing model needs to be retrained only in special cases. In fact, the workload of manual operation is very small.

Step D, the chain code tracking is post-processed.

Since the original captured image often has a large number of noise points, there are still many environmental noises that are difficult to eliminate in the curve extraction results of the above steps, and most of them are introduced by color deviation resulting in classification errors. Since the present invention is mainly aimed at extracting a relatively thin hydrological curve image, if the noise is removed by a common erosion or filtering method, the curve will often become very thin or even broken again, and satisfactory results cannot be obtained. The ideal goal is to remove the noise points without any change in the hydrological curve. In order to achieve this goal, the chain code tracking method can be used for post-processing.

The chain code tracking method can track and record the information of each connected domain in the graph in a chain code manner, that is, mark each connected domain for each pixel, and record the size and border position of each connected domain. The size of the connected domain is not based on the number of pixels, but the area of its smallest circumscribed rectangle.

For the result image after classification and extraction, track the pixel points predicted as "target curve" to obtain its connected domain information; that is, binarize the image with "target curve/non-target curve", and perform the above-mentioned steps on it. Chaincode tracking. The tracking results will include the real hydrological curve target area and the noise point area. The former is usually located in a large connected domain, while the latter is relatively small. Then a certain threshold can be set on the size of the connected domain, thereby excluding those connected domains that are smaller and where the noise is located.

The reason why the maximum connected domain is not directly obtained is to prevent the curve in the image obtained after the model classification and extraction from still having broken lines. This kind of disconnection is often subtle and easy to solve. After removing the noise points, the target curve can be expanded several times.

In order to improve the processing effect, it is also possible to perform a "grid line tracing" on the image copy before the above "curve tracing" process to determine the area where the grid lines are located, and perform the above "curve tracing" in this area. That is, the image copy is binarized with "grid line/non-grid line", and chain code tracking is performed on it, and the border position of the maximum connected domain is obtained as the outer edge of the grid line. The purpose of this operation is to remove all pixels outside the grid lines that are not related to curve extraction, reducing the complexity of curve tracing.

In order to verify the effect of the present invention, a plurality of color hydrological data images are selected for experiments, and the above-mentioned classification and extraction process is performed on them. It is agreed that the size of the selected window is 7*7, and the feature combination of the sampled points is the RGB color value, HSI color value and Lab color value of each pixel. A total of 9 feature values; that is, the dimensions of the samples involved are all 7 *7*9=441. And it is agreed that the selected machine learning method is support vector machine SVM. Taking Figure 2 as an example, the training sample set is initially empty. First, initial sampling is performed on Figure 2. After obtaining a sufficient number of training samples, the SVM classifier is trained and used to classify and predict Figure 2. The result is shown in Figure 5a. The black points are the points whose prediction result is "hydrological curve", and the gray points are "grid lines". It can be seen that the classification effect of Fig. 2 is not satisfactory at this time, and there are many broken lines, especially two large broken lines. Re-sampling these broken lines several times, after several rounds of retraining, the effect of the generated SVM classifier is improved, and the broken lines of the classification results in Figure 2 are resolved, as shown in Figure 5b.

The current SVM classifier is used to try to classify and predict Figure 3, and the result is shown in Figure 6a. The curve obtained at this time has no broken line phenomenon, but there is too much noise, and the classification effect is not good. These noise points are sampled again, a new round of classifier is trained, and the result of classifying Figure 3 again is shown in Figure 6b. At this time, the classification effect is better, and it is considered that the current classifier has been able to complete the classification requirements for these two images. This can be repeated for other images if desired.

Continue post-processing on the classification results of Figure 5b and Figure 6b, remove the unnecessary "grid line" gray points, use chain code tracking to eliminate noise points, and perform several expansion operations on the curve to obtain the curve extraction. The results are shown in Figure 7a and Figure 7b. It can be seen that the present invention has completed the extraction of the finer hydrological curves in FIGS. 2 and 3 .

The hydrological curve extraction method based on the machine learning method of the present invention is based on the classification of the sample mode, as long as sufficient training samples are selected, the influence of lighting and other issues does not need to be considered; using offline learning, it is not necessary to re-collect each image Sample training; incrementally select and add training samples to accommodate incoming new classification requirements. The invention can better solve the problem of wire breakage which is easily generated when the thin wire is extracted, and has good research value.

Claims (9)

1. a hydrological curve extraction method based on machine learning, is characterized in that, comprises the following steps: Step A, selecting the scale of the sampling window and the target feature to be sampled, and collecting a representative training sample set accordingly; Step B, using the machine learning method to train and generate a classification prediction model from the training samples; In step C, each pixel in the image to be processed is collected according to the sampling window to obtain the target feature as the sample to be classified, and the classification prediction model is used for classification; Step D, judge whether the classification result of each pixel in the image to be processed reaches the expectation, whether the curve is extracted completely and whether there are other regions with obvious classification errors; if the classification result reaches the expectation, then go to step F; otherwise, go to step E; Step E. Select representative sample points from the broken line area of the curve and the area with obvious classification errors, add them to the training sample set after sampling, and repeat steps B-D; Step F, post-processing the processed image. 2 . The method for extracting hydrological curves based on machine learning according to claim 1 , wherein in step A, the scale of the window is scalable. 3 . 3. The method for extracting hydrological curves based on machine learning as claimed in claim 1, wherein in step A, the selection of local features adopts the combination of various different types of features, and the features include color features, gradient features, Texture features and SIFT features. 4 . The method for extracting hydrological curves based on machine learning according to claim 1 , wherein, in step B, the machine learning method comprises support vector machines, neural network methods, and combinations thereof. 5 . 5. the hydrological curve extraction method based on machine learning as claimed in claim 1, it is characterized in that, in step C, traverse all the pixel points that can use the current sampling window to extract feature, obtain corresponding each local feature value according to the window and form to be Categorical sample vector. 6. the hydrological curve extraction method based on machine learning as claimed in claim 1 is characterized in that, step E is further: the sample point selected should be added to the original training set in incremental form as training sample and retrain to generate prediction model to make directional adjustments to the original model. 7 . The method for extracting hydrological curves based on machine learning according to claim 1 , wherein in step F, the post-processing method is a combination of chain code tracking and dilation processing. 8 . 8 . The method for extracting hydrological curves based on machine learning according to claim 1 , wherein a sample category “grid line” needs to be added, so as to locate the drawing area on the graph when processing the image. 9 . 9. The method for extracting hydrological curves based on machine learning as claimed in claim 7, wherein the chain code tracking is used to track the connected domains existing in the image, and calculate and record the size of these connected domains, thereby excluding the A connected domain of smaller size consisting of noise.