Remote sensing image object recognition method and system based on deep learning

Abstract

The present invention discloses a remote sensing image object recognition method and system based on deep learning, which relates to the field of image recognition technology, including: obtaining the imaging ratio of the remote sensing image; obtaining at least one recognition area; obtaining at least one irregular object classification and at least one regular object classification; obtaining at least one pre-recognition feature; forming a specific recognition feature of the irregular object classification; using the regular object classification to recognize the object to be recognized; when the regular object classification cannot be used for recognition, using the irregular object classification to recognize the object to be recognized; when the irregular object classification cannot be used for recognition, performing pixel comparison between the object to be recognized and the suspicious sample object to obtain the final sample object. By setting an object classification module, a feature formation module, a recognition judgment module and a comprehensive comparison module, the recognition efficiency can be guaranteed while the integrity of the recognition can be guaranteed, and the situation where individual objects cannot be recognized will not occur.

Description

Technical Field

[0001] The present invention relates to the field of image recognition technology, and in particular to a remote sensing image object recognition method and system based on deep learning. Background Art

[0002] With the advancement of science and technology and the advent of the information age, the application of remote sensing data processing and image recognition technology is becoming more and more extensive. Remote sensing data processing and image recognition technology can extract a large amount of geographic information, providing an effective means for research and decision-making in industries such as geology, environment, and agriculture.

[0003] Since there are various irregular images in remote sensing images, accurate identification requires comprehensive identification of the pixels in the image, which results in a huge amount of calculation. When multiple objects are identified simultaneously, a more obvious delay will occur, and the efficiency of recognition needs to be improved. Summary of the invention

[0004] In order to solve the above technical problems, a remote sensing image object recognition method and system based on deep learning are provided. This technical solution solves the problem raised in the above background technology that due to the presence of various irregular images in remote sensing images, accurate recognition requires comprehensive recognition of the pixels in the image, which results in a huge amount of calculation. When multiple objects are recognized simultaneously, a more obvious delay will occur, and the recognition efficiency needs to be improved.

[0005] In order to achieve the above purpose, the technical solution adopted by the present invention is:

[0006] Remote sensing image object recognition methods based on deep learning include:

[0007] Obtain the imaging ratio of remote sensing images;

[0008] Dividing the area of ​​the remote sensing image to obtain at least one identification area;

[0009] At least one sample object and its corresponding sample image are obtained, the sample object includes the types of all objects in the remote sensing image, the name of the sample object is known, and based on the sample image, the sample object is classified to obtain at least one irregular object classification and at least one regular object classification, wherein the irregular objects in the irregular object classification are objects of the same type, and the regular objects in the regular object classification are objects of the same type;

[0010] Performing feature formation on the irregular object in the irregular object classification to obtain at least one pre-recognition feature, wherein the at least one pre-recognition feature is paired with the irregular object classification;

[0011] Based on the pre-recognition features, a specific recognition feature for classifying irregular objects is formed;

[0012] Obtain an image of the object to be identified, and use rule-based object classification to identify the object to be identified;

[0013] When regular object classification cannot be used to identify an object, irregular object classification is used to identify the object to be identified;

[0014] When the irregular object classification fails to identify the object, the actual size of the object to be identified is obtained based on the imaging ratio of the remote sensing image, the identification area where the object to be identified is located is obtained as the target identification area, and based on the actual size of the object to be identified and the target identification area, at least one sample object close to the object to be identified is obtained as a suspicious sample object;

[0015] Perform pixel comparison on the object to be identified and at least one suspicious sample object to obtain a final sample object.

[0016] Preferably, dividing the area of ​​the remote sensing image to obtain at least one identification area comprises the following steps:

[0017] Performing attribute classification of the region, wherein the attribute classification includes mountain, plain, forest and city;

[0018] The region of the remote sensing image is divided according to the attribute classification of the region to obtain at least one identification region, wherein the attribute classification of the regions within the identification region is consistent.

[0019] Preferably, the classifying of the sample objects to obtain at least one irregular object classification and at least one regular object classification comprises the following steps:

[0020] Aggregate sample objects with the same name into a sample object set;

[0021] Perform contour recognition on sample images of sample objects in the sample object set to obtain at least one sample contour, and determine whether the at least one sample contour can obtain the same contour through scaling and rotation. If so, aggregate the sample objects in the sample object set to form a regular object classification; if not, aggregate the sample objects in the sample object set to form an irregular object classification.

[0022] Preferably, the step of forming features of the irregular objects in the irregular object classification to obtain at least one pre-recognition feature comprises the following steps:

[0023] Acquire a sample data set, where the sample data set consists of images;

[0024] Use deep learning to build a preliminary neural network model, and use the sample data set to train the preliminary neural network model;

[0025] During the training process, based on the training results, the input layer and the output layer of the preliminary neural network model are adjusted until the output result of the preliminary neural network model meets the preset requirements, wherein the preset requirements are that the features output by the preliminary neural network model can distinguish the images in the sample data set;

[0026] Based on the preliminary neural network model, feature formation is performed on the irregular object to obtain at least one preliminary recognition feature;

[0027] Verifying at least one preliminary identification feature, and specifying that an object identified by at least one preliminary identification feature of an irregular object classification belongs to the irregular object classification;

[0028] Determine whether there is an object that belongs to two of the irregular object categories at the same time. If not, the verification is completed. If so, use the neural network preliminary model to supplement the preliminary identification features of the two irregular object categories to which the object belongs at the same time, until the object belongs to only one irregular object category.

[0029] At least one preliminary recognition feature for classifying the irregular object is used as at least one pre-recognition feature.

[0030] Preferably, the forming of specific identification features for classifying irregular objects based on pre-identification features comprises the following steps:

[0031] The irregular object classification is compared with at least one pre-recognition feature of the other irregular object classifications to obtain a specific recognition feature, wherein the specific recognition feature is the pre-recognition feature of the irregular object classification, and the specific recognition feature is different from at least one pre-recognition feature of the other irregular object classifications.

[0032] Preferably, the use of rule object classification to identify the object to be identified includes the following steps:

[0033] Obtain the image size range of the remote sensing image, and determine the magnification range based on the image size range;

[0034] The magnification range is divided at equal intervals to obtain at least one magnification point;

[0035] Divide the rotation range at equal intervals to obtain at least one rotation point, wherein the rotation range is from 0 to 360 degrees;

[0036] At least one magnification point is combined with at least one rotation point to form at least one transformation group;

[0037] Using the parameters of the transformation group to transform the image of the object to be identified, obtaining at least one transformed image;

[0038] When there is a sample image of a rule object in the rule object classification that is consistent with one of the transformed images, the name of the rule object in the rule object classification is used as the name of the object to be identified, otherwise, the rule object classification cannot be used for identification.

[0039] Preferably, the use of irregular object classification to identify the object to be identified includes the following steps:

[0040] Using the specific identification features of the irregular object classification to identify the object to be identified, and obtaining at least one feature to be identified;

[0041] Identifying irregular objects classified by the irregular object classification using the specific identification feature to obtain at least one sample identification feature;

[0042] Acquire an irregular object corresponding to at least one sample identification feature that is consistent with at least one feature to be identified as a feature irregular object;

[0043] The name corresponding to the feature irregular object is used as the name of the object to be identified;

[0044] When there is no characteristic irregular object, it cannot be identified using irregular object classification.

[0045] Preferably, the step of obtaining at least one sample object close to the object to be identified as a suspicious sample object comprises the following steps:

[0046] Acquire at least one sample object whose size difference from the actual size is smaller than a preset value as a preliminary sample object;

[0047] Obtain preliminary sample objects of the same type as the target recognition area in the recognition area as suspicious sample objects.

[0048] Preferably, performing pixel comparison on the object to be identified and at least one suspicious sample object to obtain the final sample object comprises the following steps:

[0049] Taking the average of the pixel values ​​of the pixel points of the image of the object to be identified, obtaining the average value of the pixel value to be identified;

[0050] Taking the average of the pixel values ​​of the pixel points of the sample image of the suspicious sample object to obtain the average of the suspicious pixel values;

[0051] Obtaining suspicious sample objects corresponding to the suspicious pixel value mean whose difference from the pixel value mean to be identified is less than the preset pixel value as the preliminary sample objects;

[0052] Obtaining a value range of pixel values ​​of the sample image, and evenly dividing the value range of the pixel values ​​to obtain at least one pixel value interval;

[0053] Dividing the prepared sample object into pixels to obtain at least one prepared pixel point set, wherein the pixels in the prepared pixel point set belong to the same pixel value interval;

[0054] Calculating the proportion of the prepared pixel point set in at least one prepared pixel point set as the prepared proportion;

[0055] Dividing the object to be identified into pixels to obtain at least one set of pixels to be identified, wherein the pixels in the set of pixels to be identified belong to the same pixel value interval;

[0056] Calculate the proportion of the to-be-identified pixel point set in at least one to-be-identified pixel point set as the to-be-identified proportion;

[0057] Based on the proportion to be identified and the prepared proportion, a final sample object is obtained, wherein the final sample object is one of the prepared sample objects, and the difference between at least one prepared proportion distribution of the final sample object and at least one to-be-identified proportion distribution of the object to be identified is the smallest.

[0058] A remote sensing image object recognition system based on deep learning is used to implement the above remote sensing image object recognition method based on deep learning, including:

[0059] A ratio acquisition module, wherein the ratio acquisition module acquires the imaging ratio of the remote sensing image;

[0060] A region division module, wherein the region division module divides the region of the remote sensing image to obtain at least one identification region;

[0061] An object classification module, wherein the object classification module obtains at least one sample object and its corresponding sample image, wherein the sample object includes the types of all objects in the remote sensing image, and the names of the sample objects are known. Based on the sample image, the sample objects are classified to obtain at least one irregular object classification and at least one regular object classification;

[0062] A feature formation module, wherein the feature formation module performs feature formation on the irregular object in the irregular object classification to obtain at least one pre-recognition feature, wherein the at least one pre-recognition feature is paired with the irregular object classification to form a specific recognition feature of the irregular object classification based on the pre-recognition feature;

[0063] an identification and judgment module, wherein the identification and judgment module obtains an image of an object to be identified, uses regular object classification to identify the object to be identified, and when the regular object classification cannot be used to identify the object, uses irregular object classification to identify the object to be identified, and when the irregular object classification cannot be used to identify the object, obtains the actual size of the object to be identified based on the imaging ratio of the remote sensing image, obtains the identification area where the object to be identified is located as the target identification area, and obtains at least one sample object close to the object to be identified as a suspicious sample object based on the actual size of the object to be identified and the target identification area;

[0064] A comprehensive comparison module performs pixel comparison on the object to be identified and at least one suspicious sample object to obtain a final sample object.

[0065] Compared with the prior art, the present invention has the following beneficial effects:

[0066] By setting up an object classification module, a feature formation module, an identification and judgment module, and a comprehensive comparison module, it is possible to perform feature recognition on objects and classify and identify them according to their attributes. Therefore, when identifying, there is no need to retrieve a comprehensive image, and only the features generated by deep learning need to be used for recognition. This can greatly reduce the amount of data for comparison, thereby speeding up the recognition speed and improving the recognition efficiency. When rapid recognition is not possible, the image is fully recognized. Considering that the image cannot be recognized by the common means of contour lights, the distribution of its pixel values ​​is considered. The pixel value distributions of the same type of objects have a high degree of similarity. Therefore, the final recognition can be completed, thereby ensuring the recognition efficiency while also ensuring the integrity of the recognition, and there will be no situation where individual objects cannot be recognized. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 A schematic diagram of the process of the remote sensing image object recognition method based on deep learning of the present invention;

[0068] Figure 2 A schematic diagram of the process of dividing the area of ​​a remote sensing image to obtain at least one identification area according to the present invention;

[0069] Figure 3 A schematic diagram of a flow chart of classifying sample objects to obtain at least one irregular object classification and at least one regular object classification according to the present invention;

[0070] Figure 4 A schematic diagram of a process of forming features of irregular objects in the classification of irregular objects to obtain at least one pre-identification feature according to the present invention;

[0071] Figure 5 A schematic diagram of a process of classifying objects using rules of the present invention and identifying objects to be identified;

[0072] Figure 6 A schematic diagram of a process of identifying an object to be identified by using irregular object classification according to the present invention;

[0073] Figure 7 A schematic diagram of a process of obtaining at least one sample object close to an object to be identified as a suspicious sample object according to the present invention;

[0074] Figure 8 This is a schematic diagram of the process of performing pixel comparison between an object to be identified and at least one suspicious sample object to obtain a final sample object according to the present invention. DETAILED DESCRIPTION

[0075] The following description is used to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are only examples, and those skilled in the art may think of other obvious variations.

[0076] Reference Figure 1 As shown, the remote sensing image object recognition method based on deep learning includes:

[0077] Obtain the imaging ratio of remote sensing images;

[0078] Dividing the area of ​​the remote sensing image to obtain at least one identification area;

[0079] At least one sample object and its corresponding sample image are obtained, the sample object includes the types of all objects in the remote sensing image, the name of the sample object is known, and based on the sample image, the sample object is classified to obtain at least one irregular object classification and at least one regular object classification, wherein the irregular objects in the irregular object classification are objects of the same type, and the regular objects in the regular object classification are objects of the same type;

[0080] Performing feature formation on the irregular object in the irregular object classification to obtain at least one pre-recognition feature, wherein the at least one pre-recognition feature is paired with the irregular object classification;

[0081] Based on the pre-recognition features, a specific recognition feature for classifying irregular objects is formed;

[0082] Obtain an image of the object to be identified, and use rule-based object classification to identify the object to be identified;

[0083] When regular object classification cannot be used to identify an object, irregular object classification is used to identify the object to be identified;

[0084] When the irregular object classification fails to identify the object, the actual size of the object to be identified is obtained based on the imaging ratio of the remote sensing image, the identification area where the object to be identified is located is obtained as the target identification area, and based on the actual size of the object to be identified and the target identification area, at least one sample object close to the object to be identified is obtained as a suspicious sample object;

[0085] Perform pixel comparison on the object to be identified and at least one suspicious sample object to obtain a final sample object.

[0086] Reference Figure 2 As shown, dividing the area of ​​the remote sensing image to obtain at least one identification area includes the following steps:

[0087] Performing attribute classification of the region, wherein the attribute classification includes mountain, plain, forest and city;

[0088] The region of the remote sensing image is divided according to the attribute classification of the region to obtain at least one identification region, wherein the attribute classification of the regions within the identification region is consistent.

[0089] The purpose of dividing the remote sensing images into regions is to reduce the number of images involved in a comprehensive comparison when irregular object classification and regular object classification cannot be used for comparison, thereby limiting the amount of data for the comprehensive comparison, thereby compressing the comparison and recognition and improving the recognition efficiency.

[0090] Reference Figure 3 As shown, classifying the sample objects to obtain at least one irregular object classification and at least one regular object classification includes the following steps:

[0091] Aggregate sample objects with the same name into a sample object set;

[0092] Perform contour recognition on sample images of sample objects in the sample object set to obtain at least one sample contour, and determine whether the at least one sample contour can obtain the same contour through scaling and rotation. If so, aggregate the sample objects in the sample object set to form a regular object classification; if not, aggregate the sample objects in the sample object set to form an irregular object classification.

[0093] In the recognition process, there are objects that are easy to recognize and objects that are not easy to recognize. The shapes of the objects that are easy to recognize will not change, and there are only differences in size and shooting angle. For these objects, only contour recognition is required, and there is no need to search inside the image. Therefore, they need to be classified. In this way, the recognition time can be reduced during recognition. At the same time, there are objects that are not easy to recognize. The shapes of the same type of difficult to recognize objects are quite different, and they cannot be recognized using contours. However, they will have different features from objects of different types, and the features can be recognized and used as the basis for identifying the difficult to recognize objects. Similarly, there is no need to fully recognize the difficult to recognize objects, only to recognize their corresponding features, which can also limit the amount of data to be recognized and reduce the recognition time.

[0094] Reference Figure 4 As shown, performing feature formation on the irregular object in the irregular object classification to obtain at least one pre-recognition feature includes the following steps:

[0095] Acquire a sample data set, where the sample data set consists of images;

[0096] Use deep learning to build a preliminary neural network model, and use the sample data set to train the preliminary neural network model;

[0097] During the training process, based on the training results, the input layer and the output layer of the preliminary neural network model are adjusted until the output result of the preliminary neural network model meets the preset requirements, wherein the preset requirements are that the features output by the preliminary neural network model can distinguish the images in the sample data set;

[0098] Based on the preliminary neural network model, feature formation is performed on the irregular object to obtain at least one preliminary recognition feature;

[0099] Verifying at least one preliminary identification feature, and specifying that an object identified by at least one preliminary identification feature of an irregular object classification belongs to the irregular object classification;

[0100] Determine whether there is an object that belongs to two of the irregular object categories at the same time. If not, the verification is completed. If so, use the neural network preliminary model to supplement the preliminary identification features of the two irregular object categories to which the object belongs at the same time, until the object belongs to only one irregular object category.

[0101] At least one preliminary recognition feature for classifying the irregular object is used as at least one pre-recognition feature.

[0102] The pre-recognition feature is a feature used to identify irregular objects, and may contain redundant features. Using the pre-recognition feature for identification will result in useless comparison and increase the comparison time. Therefore, the pre-recognition feature should be simplified. When simplifying, the pre-recognition feature of the irregular object classification should be compared with the pre-recognition features of other irregular object classifications to obtain a specific recognition feature of the difference. The specific recognition feature can be used to distinguish the irregular object classifications and determine the irregular object classification to which the object to be identified belongs. The objects in the irregular object classification are all of the same type. Once the irregular object classification to which the object to be identified belongs is determined, the name of the object to be identified can be determined, and the identification is completed.

[0103] Based on the pre-recognition features, forming a specific recognition feature for classifying irregular objects includes the following steps:

[0104] The irregular object classification is compared with at least one pre-recognition feature of the other irregular object classifications to obtain a specific recognition feature, wherein the specific recognition feature is the pre-recognition feature of the irregular object classification, and the specific recognition feature is different from at least one pre-recognition feature of the other irregular object classifications.

[0105] Reference Figure 5 As shown, using rule object classification, identifying the object to be identified includes the following steps:

[0106] Obtain the image size range of the remote sensing image, and determine the magnification range based on the image size range;

[0107] The magnification range is divided at equal intervals to obtain at least one magnification point;

[0108] Divide the rotation range at equal intervals to obtain at least one rotation point, wherein the rotation range is from 0 to 360 degrees;

[0109] At least one magnification point is combined with at least one rotation point to form at least one transformation group;

[0110] Using the parameters of the transformation group to transform the image of the object to be identified, obtaining at least one transformed image;

[0111] When there is a sample image of a rule object in the rule object classification that is consistent with one of the transformed images, the name of the rule object in the rule object classification is used as the name of the object to be identified, otherwise, the rule object classification cannot be used for identification.

[0112] When using rule-based object classification for classification, it is necessary to rotate and scale the object to be identified. Therefore, a transformation group is formed. The transformation group can be approximated to any degree, which is equivalent to including all possibilities of rotation and scaling of the object to be identified. Therefore, identification can be performed based on whether the sample image is consistent with one of the transformed images. The consistency here is actually an approximate description. When the difference is less than the allowable error, it is called consistency.

[0113] Reference Figure 6 As shown, using irregular object classification, identifying the object to be identified includes the following steps:

[0114] Using the specific identification features of the irregular object classification to identify the object to be identified, and obtaining at least one feature to be identified;

[0115] Identifying irregular objects classified by the irregular object classification using the specific identification feature to obtain at least one sample identification feature;

[0116] Acquire an irregular object corresponding to at least one sample identification feature that is consistent with at least one feature to be identified as a feature irregular object;

[0117] The name corresponding to the feature irregular object is used as the name of the object to be identified;

[0118] When there is no characteristic irregular object, it cannot be identified using irregular object classification.

[0119] Reference Figure 7 As shown, obtaining at least one sample object close to the object to be identified as a suspicious sample object includes the following steps:

[0120] Acquire at least one sample object whose size difference from the actual size is smaller than a preset value as a preliminary sample object;

[0121] Obtain preliminary sample objects of the same type as the target recognition area in the recognition area as suspicious sample objects.

[0122] Reference Figure 8 As shown, performing pixel comparison between the object to be identified and at least one suspicious sample object to obtain the final sample object includes the following steps:

[0123] Taking the average of the pixel values ​​of the pixel points of the image of the object to be identified, obtaining the average value of the pixel value to be identified;

[0124] Taking the average of the pixel values ​​of the pixel points of the sample image of the suspicious sample object to obtain the average of the suspicious pixel values;

[0125] Obtaining suspicious sample objects corresponding to the suspicious pixel value mean whose difference from the pixel value mean to be identified is less than the preset pixel value as the preliminary sample objects;

[0126] Obtaining a value range of pixel values ​​of the sample image, and evenly dividing the value range of the pixel values ​​to obtain at least one pixel value interval;

[0127] Dividing the prepared sample object into pixels to obtain at least one prepared pixel point set, wherein the pixels in the prepared pixel point set belong to the same pixel value interval;

[0128] Calculating the proportion of the prepared pixel point set in at least one prepared pixel point set as the prepared proportion;

[0129] Dividing the object to be identified into pixels to obtain at least one set of pixels to be identified, wherein the pixels in the set of pixels to be identified belong to the same pixel value interval;

[0130] Calculate the proportion of the to-be-identified pixel point set in at least one to-be-identified pixel point set as the to-be-identified proportion;

[0131] Based on the proportion to be identified and the prepared proportion, a final sample object is obtained, wherein the final sample object is one of the prepared sample objects, and the difference between at least one prepared proportion distribution of the final sample object and at least one to-be-identified proportion distribution of the object to be identified is the smallest.

[0132] When regular object classification and irregular object classification are both unrecognizable, it means that the object to be identified is not a regular object and its shape cannot be used as a reference for identification. However, since the pixel value distributions of objects of the same type are similar and the pixel value distributions of objects of different types are quite different, the pixel value distributions are statistically analyzed to complete the identification. However, the identification here involves a comprehensive image retrieval, which involves a large amount of data and will result in a scientifically perceived delay.

[0133] A remote sensing image object recognition system based on deep learning is used to implement the above remote sensing image object recognition method based on deep learning, including:

[0134] A ratio acquisition module, wherein the ratio acquisition module acquires the imaging ratio of the remote sensing image;

[0135] A region division module, wherein the region division module divides the region of the remote sensing image to obtain at least one identification region;

[0136] An object classification module, wherein the object classification module obtains at least one sample object and its corresponding sample image, wherein the sample object includes the types of all objects in the remote sensing image, and the names of the sample objects are known. Based on the sample image, the sample objects are classified to obtain at least one irregular object classification and at least one regular object classification;

[0137] A feature formation module, wherein the feature formation module performs feature formation on the irregular object in the irregular object classification to obtain at least one pre-recognition feature, wherein the at least one pre-recognition feature is paired with the irregular object classification to form a specific recognition feature of the irregular object classification based on the pre-recognition feature;

[0138] an identification and judgment module, wherein the identification and judgment module obtains an image of an object to be identified, uses regular object classification to identify the object to be identified, and when the regular object classification cannot be used to identify the object, uses irregular object classification to identify the object to be identified, and when the irregular object classification cannot be used to identify the object, obtains the actual size of the object to be identified based on the imaging ratio of the remote sensing image, obtains the identification area where the object to be identified is located as the target identification area, and obtains at least one sample object close to the object to be identified as a suspicious sample object based on the actual size of the object to be identified and the target identification area;

[0139] A comprehensive comparison module performs pixel comparison on the object to be identified and at least one suspicious sample object to obtain a final sample object.

[0140] Furthermore, the present solution also proposes a storage medium on which a computer-readable program is stored, and when the computer-readable program is called, the above-mentioned remote sensing image object recognition method based on deep learning is executed.

[0141] It is understandable that the storage medium may be a magnetic medium, such as a floppy disk, a hard disk, or a magnetic tape; an optical medium, such as a DVD; or a semiconductor medium, such as a solid state drive (SSD).

[0142] To sum up, the advantages of the present invention are: by setting an object classification module, a feature formation module, an identification and judgment module and a comprehensive comparison module, the features of the object can be recognized, and the classified recognition can be performed according to the attributes of the object, so that when recognizing, there is no need to retrieve a comprehensive image, and only the features generated by deep learning need to be used for recognition, which can greatly reduce the amount of data for comparison, thereby speeding up the recognition speed and improving the recognition efficiency. When rapid recognition is not possible, the image is comprehensively recognized. Considering that the image cannot be recognized by the common means of contour lights, the distribution of its pixel values ​​is considered. The pixel value distributions of the same type of objects have a high degree of similarity. Therefore, the final recognition can be completed, thereby ensuring the recognition efficiency while also ensuring the integrity of the recognition, and there will be no situation where individual objects cannot be recognized.

[0143] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions only describe the principles of the present invention. The present invention may be subject to various changes and improvements without departing from the spirit and scope of the present invention. These changes and improvements fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the attached claims and their equivalents.

Claims

1. A remote sensing image object recognition method based on deep learning, characterized in that: include: Obtain the imaging ratio of remote sensing images; Dividing the area of ​​the remote sensing image to obtain at least one identification area; At least one sample object and its corresponding sample image are obtained, the sample object includes the types of all objects in the remote sensing image, the name of the sample object is known, and based on the sample image, the sample object is classified to obtain at least one irregular object classification and at least one regular object classification, wherein the irregular objects in the irregular object classification are objects of the same type, and the regular objects in the regular object classification are objects of the same type; Performing feature formation on the irregular object in the irregular object classification to obtain at least one pre-recognition feature, wherein the at least one pre-recognition feature is paired with the irregular object classification; Based on the pre-recognition features, a specific recognition feature for classifying irregular objects is formed; Obtain an image of the object to be identified, and use rule-based object classification to identify the object to be identified; When regular object classification cannot be used to identify an object, irregular object classification is used to identify the object to be identified; When the irregular object classification fails to identify the object, the actual size of the object to be identified is obtained based on the imaging ratio of the remote sensing image, the identification area where the object to be identified is located is obtained as the target identification area, and based on the actual size of the object to be identified and the target identification area, at least one sample object close to the object to be identified is obtained as a suspicious sample object; Perform pixel comparison on the object to be identified and at least one suspicious sample object to obtain a final sample object; The step of forming features of the irregular objects in the irregular object classification to obtain at least one pre-recognition feature comprises the following steps: Acquire a sample data set, where the sample data set consists of images; Use deep learning to build a preliminary neural network model, and use sample data sets to train the preliminary neural network model; During the training process, based on the training results, the input layer and the output layer of the preliminary neural network model are adjusted until the output result of the preliminary neural network model meets the preset requirements, wherein the preset requirements are that the features output by the preliminary neural network model can distinguish the images in the sample data set; Based on the preliminary neural network model, feature formation is performed on the irregular object to obtain at least one preliminary recognition feature; Verifying at least one preliminary identification feature, and specifying that an object identified by at least one preliminary identification feature of an irregular object classification belongs to the irregular object classification; Determine whether there is an object that belongs to two of the irregular object categories at the same time. If not, the verification is completed. If so, use the neural network preliminary model to supplement the preliminary identification features of the two irregular object categories to which the object belongs at the same time, until the object belongs to only one irregular object category. using at least one preliminary recognition feature for classifying the irregular object as at least one pre-recognition feature; The forming of a specific recognition feature for classifying irregular objects based on the pre-recognition feature comprises the following steps: The irregular object classification is compared with at least one pre-recognition feature of the other irregular object classifications to obtain a specific recognition feature, wherein the specific recognition feature is the pre-recognition feature of the irregular object classification, and the specific recognition feature is different from at least one pre-recognition feature of the other irregular object classifications.

2. The remote sensing image object recognition method based on deep learning according to claim 1, characterized in that: The dividing of the area of ​​the remote sensing image to obtain at least one identification area comprises the following steps: Performing attribute classification of the region, wherein the attribute classification includes mountain, plain, forest and city; The region of the remote sensing image is divided according to the attribute classification of the region to obtain at least one identification region, wherein the attribute classification of the regions within the identification region is consistent.

3. The remote sensing image object recognition method based on deep learning according to claim 2, characterized in that: The classifying of the sample objects to obtain at least one irregular object classification and at least one regular object classification comprises the following steps: Aggregate sample objects with the same name into a sample object set; Perform contour recognition on sample images of sample objects in the sample object set to obtain at least one sample contour, and determine whether the at least one sample contour can obtain the same contour through scaling and rotation. If so, aggregate the sample objects in the sample object set to form a regular object classification; if not, aggregate the sample objects in the sample object set to form an irregular object classification.

4. The remote sensing image object recognition method based on deep learning according to claim 3 is characterized in that: The use of rule object classification to identify the object to be identified includes the following steps: Obtain the image size range of the remote sensing image, and determine the magnification range based on the image size range; The magnification range is divided at equal intervals to obtain at least one magnification point; Divide the rotation range at equal intervals to obtain at least one rotation point, wherein the rotation range is from 0 to 360 degrees; At least one magnification point is combined with at least one rotation point to form at least one transformation group; Using the parameters of the transformation group to transform the image of the object to be identified, obtaining at least one transformed image; When there is a sample image of a rule object in the rule object classification that is consistent with one of the transformed images, the name of the rule object in the rule object classification is used as the name of the object to be identified, otherwise, the rule object classification cannot be used for identification.

5. The remote sensing image object recognition method based on deep learning according to claim 4 is characterized in that: The method of using irregular object classification to identify the object to be identified includes the following steps: Using the specific identification features of the irregular object classification to identify the object to be identified, and obtaining at least one feature to be identified; Identifying irregular objects classified by the irregular object classification using the specific identification feature to obtain at least one sample identification feature; Acquire an irregular object corresponding to at least one sample identification feature that is consistent with at least one feature to be identified as a feature irregular object; The name corresponding to the feature irregular object is used as the name of the object to be identified; When there is no characteristic irregular object, it cannot be identified using irregular object classification.

6. The remote sensing image object recognition method based on deep learning according to claim 5, characterized in that: The step of obtaining at least one sample object close to the object to be identified as a suspicious sample object comprises the following steps: Acquire at least one sample object whose size difference from the actual size is smaller than a preset value as a preliminary sample object; Obtain preliminary sample objects of the same type as the target recognition area in the recognition area as suspicious sample objects.

7. The remote sensing image object recognition method based on deep learning according to claim 6, characterized in that: The pixel comparison between the object to be identified and at least one suspicious sample object to obtain the final sample object comprises the following steps: Taking the average of the pixel values ​​of the pixel points of the image of the object to be identified, obtaining the average value of the pixel value to be identified; Taking the average of the pixel values ​​of the pixel points of the sample image of the suspicious sample object to obtain the average of the suspicious pixel values; Obtaining suspicious sample objects corresponding to the suspicious pixel value mean whose difference from the pixel value mean to be identified is less than the preset pixel value as the preliminary sample objects; Obtaining a value range of pixel values ​​of the sample image, and evenly dividing the value range of the pixel values ​​to obtain at least one pixel value interval; Dividing the prepared sample object into pixels to obtain at least one prepared pixel point set, wherein the pixels in the prepared pixel point set belong to the same pixel value interval; Calculating the proportion of the prepared pixel point set in at least one prepared pixel point set as the prepared proportion; Dividing the object to be identified into pixels to obtain at least one set of pixels to be identified, wherein the pixels in the set of pixels to be identified belong to the same pixel value interval; Calculate the proportion of the to-be-identified pixel point set in at least one to-be-identified pixel point set as the to-be-identified proportion; Based on the proportion to be identified and the prepared proportion, a final sample object is obtained, wherein the final sample object is one of the prepared sample objects, and the difference between at least one prepared proportion distribution of the final sample object and at least one to-be-identified proportion distribution of the object to be identified is the smallest.

8. A remote sensing image object recognition system based on deep learning, used to implement the remote sensing image object recognition method based on deep learning as described in any one of claims 1 to 7, characterized in that: include: A ratio acquisition module, wherein the ratio acquisition module acquires the imaging ratio of the remote sensing image; A region division module, wherein the region division module divides the region of the remote sensing image to obtain at least one identification region; An object classification module, wherein the object classification module obtains at least one sample object and its corresponding sample image, wherein the sample object includes the types of all objects in the remote sensing image, and the names of the sample objects are known. Based on the sample image, the sample objects are classified to obtain at least one irregular object classification and at least one regular object classification; A feature formation module, wherein the feature formation module performs feature formation on the irregular object in the irregular object classification to obtain at least one pre-recognition feature, wherein the at least one pre-recognition feature is paired with the irregular object classification to form a specific recognition feature of the irregular object classification based on the pre-recognition feature; an identification and judgment module, wherein the identification and judgment module obtains an image of an object to be identified, uses regular object classification to identify the object to be identified, and when the regular object classification cannot be used to identify the object, uses irregular object classification to identify the object to be identified, and when the irregular object classification cannot be used to identify the object, obtains the actual size of the object to be identified based on the imaging ratio of the remote sensing image, obtains the identification area where the object to be identified is located as the target identification area, and obtains at least one sample object close to the object to be identified as a suspicious sample object based on the actual size of the object to be identified and the target identification area; A comprehensive comparison module performs pixel comparison on the object to be identified and at least one suspicious sample object to obtain a final sample object.

Images