A map image object detection method considering spatial relations and co-occurrence probability

Abstract

The map image target detection method of the present invention, which takes into account spatial relationships and co-occurrence probabilities, detects multiple map targets through a basic target detection model, and on this basis describes the spatial attributes between different targets through area relationships and orientation relationships, and then improves the map target detection accuracy through the co-occurrence probability between different targets. The introduction of spatial relationships between multiple targets in this method can provide effective technical means and references for effective target extraction for intelligent recognition of large-scale map content.

Description

Technical Field

[0001] The present invention relates to the field of target detection, and in particular to a map image target detection method taking into account spatial relationships and co-occurrence probabilities. Background Art

[0002] Maps are recognized as one of the three major universal languages ​​in the world. As an important carrier of geographic information, they occupy a very important position in the expression of objective geographic space and the reconstruction of the nonlinear geographical world.

[0003] On the one hand, the map reading process is the process of reconstructing the real space in the mind of the map user, which is an important step for the map to play its role in transmitting and expressing geographic information. However, in the study of cartography, most of the objective world is mapped onto the map, and there is a lack of in-depth research on the mapping from map images to the objective world. Map making technology has gone through the stages of traditional manual mapping, digital mapping, and the current intelligent mapping. The proposed cartographic information transmission model combines ICT, artificial intelligence, the Internet, big data and other technologies. Scholars have conducted in-depth research on the cartographer stage of the model, but there is still insufficient research on the map reader stage. Therefore, in the context of the era of networking and intelligent development, the intelligent research on the map user stage needs to be deepened.

[0004] On the other hand, automated map content extraction is the only way to mine geospatial data. The widespread use of the Internet and the strong demand for geographic information have rapidly popularized Internet geographic information services (such as WMS, WFS, etc.), but as part of the Deep Web resources, this part of data cannot be directly obtained through web page analysis. However, this type of geographic information data outputs the data in the server through map rendering. Therefore, in the face of massive map image data in the network, identifying vector symbols and target objects from them can, to a certain extent, obtain, analyze, and restore the data in the geographic information server in the deep web environment. In addition, for the recognition of multi-source map content in the network environment, the map's drawing style, the attention of key areas, the error-proneness of key areas, and other features can be extracted at a macro level to evaluate the map mapping features in the network environment. At present, the research on the extraction and recognition of network map image content is in its early stages, and the recognition method for the content in map images in this field needs to be deepened.

[0005] In existing research on map image classification or content recognition, there is no effective use of the spatial or interdependent relationships between the objects to be identified in the map; however, as an important carrier of geographic information, the content expressed by the map has a very strong spatial correlation.

[0006] Therefore, technicians in this field are committed to developing a map image target detection method that takes into account spatial relationships and co-occurrence probabilities, and identifies different target objects on the map through the area, position spatial relationships between different objects and the co-occurrence probabilities between different targets, so as to solve the shortcomings of the above-mentioned existing technologies. Summary of the invention

[0007] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is that the feature extraction and recognition of the target object in the current prior art are concentrated in the training process of the model, and the methods for improving the accuracy of the model mostly start from the training process of the model, while not considering the influence of the spatial dependency between different objects on the target detection results.

[0008] To achieve the above object, the present invention provides a map image object detection method taking into account spatial relationships and co-occurrence probabilities.

[0009] The specific steps include:

[0010] Step 1: Use web crawlers to obtain various types of map image data from the Internet environment, obtain different types of targets in all map images through manual annotation, and randomly divide the manual annotation results into training and test data sets according to a certain ratio;

[0011] Step 2: Build a Retinanet network model and use it to train the model on the training data set to obtain two target detection models for target 1 and target 2;

[0012] Step 3: For the training sample data, for the existence of two objects in the sample, generate a binary data set for each map sample, and then perform statistics on different co-occurrence probabilities in the sample;

[0013] By counting the distribution of two-tuples of two targets (target 1, target 2) in the training samples, the statistical results of the number of different types of target tuples in the training samples are obtained;

[0014]

[0015] In formula (7) and formula (8),

[0016] Indicates the possibility of the existence of target 1 when target 2 exists;

[0017] count Tai Indicates the number of all images containing target 2;

[0018] Indicates the number of map images that contain both target 1 and target 2; obtained by statistically analyzing the training samples The value represents the co-occurrence degree of the target 2 detection model to the target 1 detection model;

[0019] Indicates the possibility that the image contains only target 1 but not target 2;

[0020] count all Represents the number of all training samples;

[0021] Indicates the number of sample images that only contain target 1;

[0022] Step 4: For the training data, use formula (1) and formula (2) to count the area and relative position of the two targets, and obtain the area ratio range (range_ration_area) of the model's circumscribed rectangle and the relative direction tangent value range (range_tan_gravity) of the line connecting the centers of the two circumscribed rectangles; in the prediction results of the model for unknown map data, range_ration_area and range_tan_gravity are used as discriminant annotations for the simultaneous existence of two targets in the prediction results;

[0023] ratio_area = area_box2 / area_box1 formula (1)

[0024] tan_gravity=delt_y / delt_x formula (2)

[0025] In formula (1) and formula (2),

[0026] ratio_area is the area ratio of the circumscribed rectangle;

[0027] tan_gravity is the relative direction tangent value of the line connecting the centers of the circumscribed rectangle;

[0028] area_box1 is the area of ​​the circumscribed rectangle of target 1;

[0029] area_box2 is the area of ​​the circumscribed rectangle of target 2;

[0030] delt_y is the difference between the ordinates of the center points of the circumscribed rectangles of target 1 and target 2;

[0031] delt_x is the difference between the horizontal coordinates of the center points of the circumscribed rectangles of target 1 and target 2;

[0032] Step 5: After obtaining the spatial relationship and co-occurrence probability relationship, test the test sample and use the spatial relationship to determine whether the two targets exist in the map image at the same time; if the bounding rectangles of the two targets in the prediction result cannot simultaneously satisfy the range calculated in step 4, it is considered that the two targets do not exist at the same time; if they both satisfy the two range requirements calculated in step 4, it is considered that the two targets may exist at the same time;

[0033] Furthermore, in step 5, if the image contains two targets, for the prediction results that may exist at the same time, the prediction results of the low-precision model are corrected by the co-occurrence probability of the reliability of the prediction results of the high-precision model and the reliability of the prediction results of the low-precision model; the low-precision target detection model is corrected using the following formula:

[0034]

[0035] In formula (3), formula (4), and formula (5),

[0036] Direct prediction or prediction confidence for target 1 object;

[0037] Represents the corrected confidence of the prediction result of target 1 object brought by the co-occurrence probability;

[0038] The confidence that target 1 exists independently in the image;

[0039] Represents the final prediction confidence of the target 1 object detection model after correction;

[0040] Furthermore, in step 5, when there is only one target in the test map image, the following formula is used to correct the recognition result:

[0041]

[0042] In formula (4) and formula (6),

[0043] The confidence that target 1 exists independently in the image;

[0044] Represents the final prediction confidence of the target 1 object detection model after correction;

[0045] Direct prediction or prediction confidence for target 1 object;

[0046] By adopting the above scheme, the map image object detection method taking into account spatial relationship and co-occurrence probability disclosed by the present invention has the following advantages:

[0047] The map image target detection method that takes into account spatial relationships and co-occurrence probabilities disclosed in the present invention is aimed at the defect problem that the existing detection of different map targets in the map only uses the physical characteristics of the target itself, but does not consider the spatial relationship and co-occurrence probability between the targets. In view of the fact that the map is a carrier of geographic information, the objects it expresses have a high degree of spatial attributes. In the process of detecting different targets in the map, if the spatial attributes are not considered, a large number of associated attributes of the map targets will be lost. The present invention detects multiple map targets through a basic target detection model, and on this basis describes the spatial attributes between different targets through area relationships and azimuth relationships, thereby improving the map target detection accuracy. The introduction of spatial relationships between multiple targets in the method can provide effective technical means and references for target extraction for intelligent recognition of large-scale map content.

[0048] The concept, specific technical scheme and technical effects of the present invention will be further described below in conjunction with specific implementation methods to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a flowchart of map image object detection correction taking into account spatial relationship and co-occurrence probability of the present invention;

[0050] Figure 2 This is a statistical chart of the accuracy improvement effect of the laboratory target 1 target detection model after correction; DETAILED DESCRIPTION

[0051] The following introduces several preferred embodiments of the present invention to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, which are exemplary descriptions, and the protection scope of the present invention is not limited to the embodiments mentioned in the text.

[0052] Example 1: A map image object detection method taking into account spatial relationship and co-occurrence probability

[0053] Step 1: Use web crawlers to obtain various types of map image data from the Internet environment, obtain different types of targets in all map images through manual annotation, and randomly divide the manual annotation results into training and test data sets in a ratio of 4:1;

[0054] Step 2: Build a Retinanet network model and use it to train the model on the training data set to obtain two target detection models for target 1 and target 2;

[0055] Step 3: For the training sample data, for the existence of two objects in the sample, generate a binary data set for each map sample, and then perform statistics on different co-occurrence probabilities in the sample;

[0056] By counting the distribution of the binary pairs of two targets (target 1, target 2) in the training samples, we get the results shown in Table 1.

[0057] Table 1 Statistics of the number of different types of target bigrams in training samples

[0058]

[0059] From Table 1, we can get

[0060]

[0061] Step 4: Use formula (1) and formula (2) to calculate the spatial relationship between two targets in the map where the training sample exists simultaneously, i.e., the object exists in the binary pair (1,1);

[0062] ratio_area = area_box2 / area_box1 formula (1)

[0063] tan_gravity=delt_y / delt_x formula (2)

[0064] Through calculation, in the training data, when two targets exist at the same time, the area ratio ratio_area is in the range of [0,0.2], and the central direction description value tan_gravity is in the range of [0,3].

[0065] Step 5: After obtaining the spatial relationship and co-occurrence probability relationship, Figure 1 The test data test process shown in the figure tests the test samples and uses the spatial relationship to determine whether two targets exist in the map image at the same time. If the image contains two targets, the low-precision target detection model is corrected using the following formula:

[0066]

[0067] When there is only one target in the test map image, use the following formula to correct the recognition result:

[0068]

[0069] In this embodiment 1, the target detection neural network model is used to detect the target on the map, and the prediction confidence of the low-precision model is effectively corrected by using the prediction confidence of the high-precision model through the two processes of space and co-occurrence probability correction, thereby improving the prediction accuracy of the low-precision target detection model; the effect of the accuracy improvement is as follows Figure 2 As shown;

[0070] Thus, the map image target detection method of the present invention taking into account the spatial relationship and co-occurrence probability uses the spatial features and co-occurrence probability between multiple targets to improve the accuracy of the low-precision model by using the high-precision model; uses the area ratio relationship and the orientation relationship between different targets to describe the spatial relationship between the two map targets 1 and 2, and uses the spatial relationship as the basis for judging whether the two targets co-occur, successfully applying the spatial relationship to the correction of the map multi-target detection results; proposes a method for transferring the confidence of the prediction result, and uses the confidence of the prediction result of the high-precision model to adjust the prediction result of the low-precision model through the co-occurrence probability between different targets, thereby effectively improving the prediction result of the low-precision model;

[0071] To sum up, the technical solution of this patent detects multiple targets on the map through a basic target detection model, and on this basis describes the spatial attributes between different targets through area relationships and orientation relationships, and then improves the map target detection accuracy through the co-occurrence probability between different targets; the introduction of spatial relationships between multiple targets in this method can provide effective technical means and references for effective target extraction for intelligent recognition of large-scale map content.

[0072] The preferred specific embodiments of the present invention are described in detail above. It should be understood that ordinary technicians in the field can make many modifications and changes based on the concept of the present invention without creative work. Therefore, all technical solutions that can be obtained by technicians in the technical field based on the concept of the present invention through logical analysis, reasoning or limited experiments on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A method for detecting map image objects taking into account spatial relationships and co-occurrence probabilities, characterized in that: The steps include: Step 1: Use web crawlers to obtain various types of map image data from the Internet environment, obtain different types of targets in all map images through manual annotation, and randomly divide the manual annotation results into training and test data sets according to a certain ratio; Step 2: Build a Retinanet network model and use it to perform model training on the training data set to obtain two target detection models for target 1 and target 2. Step 3: For the training sample data, for the presence of two targets in the sample, generate a binary data set for each map sample, and then perform statistics on different co-occurrence probabilities in the sample; Step 4: For the training data, use formula (1) and formula (2) to count the area and relative position of the two targets, and obtain the area ratio range of the target's circumscribed rectangle, recorded as range_ration_area, and the relative direction tangent value range of the line connecting the centers of the two circumscribed rectangles, recorded as range_tan_gravity; in the prediction results of the Retinanet network model for unknown map data, range_ration_area and range_tan_gravity are used as discriminant annotations for the simultaneous existence of two targets in the prediction results; ratio_area=area_box2 / area_box1 formula (1) tan_gravity=delt_y / delt_x formula (2) In formula (1) and formula (2), ratio_area is the area ratio of the circumscribed rectangle; tan_gravity is the relative direction tangent value of the line connecting the centers of the circumscribed rectangle; area_box1 is the area of ​​the circumscribed rectangle of target 1; area_box2 is the area of ​​the circumscribed rectangle of target 2; delt_y is the difference between the ordinates of the center points of the circumscribed rectangles of target 1 and target 2; delt_x is the difference between the horizontal coordinates of the center points of the circumscribed rectangles of target 1 and target 2; Step 5: After obtaining the spatial relationship and co-occurrence probability relationship, test the test sample and use the spatial relationship to determine whether the two targets exist in the map image at the same time; if the bounding rectangles of the two targets in the prediction result cannot simultaneously satisfy the two ranges calculated in step 4, it is considered that the two targets do not exist at the same time; if they simultaneously satisfy the two range requirements calculated in step 4, it is considered that the two targets may exist at the same time; In step 3, By counting the distribution of two-target binary pairs in the training samples, the statistical results of the number of binary pairs of different types of targets in the training samples are obtained; In formula (7) and formula (8), Indicates the possibility of the existence of target 1 when target 2 exists; count Tai Indicates the number of all images containing target 2; Indicates the number of map images that contain both target 1 and target 2; obtained by statistically analyzing the training samples value, the The value represents the co-occurrence of the target 2 detection model with the target 1 detection model; Indicates the possibility that the image contains only target 1 but not target 2; count all Represents the number of all training samples; Indicates the number of sample images that contain only target 1.

2. The map image object detection method taking into account spatial relationships and co-occurrence probabilities as claimed in claim 1, wherein step 5 is characterized in that: If the image contains two targets, for the prediction results that may exist at the same time, the prediction results of the low-precision target detection model are corrected by the reliability of the prediction results of the high-precision target detection model and the co-occurrence probability of the reliability of the prediction results of the low-precision target detection model; the low-precision target detection model is corrected using the following formula: In formula (3), formula (4), and formula (5), The prediction confidence obtained by directly predicting target 1; Represents the corrected confidence of the prediction result of target 1 brought by the co-occurrence probability; Represents the confidence that target 1 exists independently in the image; Represents the final prediction confidence of target 1 target detection model after correction.

3. The map image object detection method taking into account spatial relationships and co-occurrence probabilities as claimed in claim 1, wherein in step 5, In step 5, when there is only one target in the test map image, the following formula is used to correct the recognition result: In formula (4) and formula (6), Represents the confidence that target 1 exists independently in the image; Represents the final prediction confidence of target 1 target detection model after correction; The prediction confidence obtained by directly predicting target 1.

Images