Unmanned aerial vehicle target detection method and system based on deep learning

Abstract

The invention discloses an unmanned aerial vehicle target detection method and system based on deep learning, and the method comprises the steps: obtaining the size features, shape features and color features of a target, carrying out the construction of indexes for the features of the target, and obtaining a shape standard database; and designing a classifier, and evaluating and repairing the classifier. By arranging a mode recognition training module, an image correction module, a display feature extraction module and a mode matching module, the finished state and the unfinished state are distinguished during detection, appearance comparison is not performed when the product is not finished, and appearance comparison is performed when the product is finished, so that whether the appearance of the product meets the requirement or not during production can be found in time, and for the product which does not meet the requirement, the production efficiency is improved. In addition, the image correction module performs stretching transformation on the target object image to obtain a target object display schematic diagram, and the ratio of the target object display schematic diagram to the actual production field is one to one, so that the size can be conveniently judged from the image.

Description

Technical Field

[0001] The present invention relates to the field of drone detection technology, and in particular to a drone target detection method and system based on deep learning. Background Art

[0002] The use of drones for photography provides a more convenient and effective management method for many application fields. Among them, target detection, that is, detecting image targets with certain characteristics from drone aerial photos, is the most widely used application. Drones also play an important role in product production.

[0003] When a product is produced, strict requirements are imposed on the dimensional parameters of its components. However, during monitoring, it is difficult to distinguish between the finished and unfinished states, and the size judgment must be made in the finished state. Therefore, it is impossible to determine whether the dimensional parameters of each component meet the requirements. Summary of the invention

[0004] In order to solve the above technical problems, a UAV target detection method and system based on deep learning are provided. The technical solution solves the problem that the products proposed in the above background technology have strict requirements on the dimensional parameters of their components during production, but it is difficult to distinguish between the finished and unfinished states during monitoring, and the size judgment must be carried out in the finished state. Therefore, it is impossible to judge whether the dimensional parameters of each component meet the requirements.

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

[0006] A UAV target detection method based on deep learning, comprising:

[0007] Obtain the size, shape and color features of the target, construct indicators based on the target features, and obtain a standard appearance database;

[0008] Design, evaluate and repair classifiers;

[0009] Acquire the target image information by shooting with a drone and convert the target image information into electrical information;

[0010] Acquire electrical information and perform image preprocessing, including signal enhancement, signal denoising, signal recovery and image normalization;

[0011] Correct the preprocessed target image, stretch and transform the target image, and obtain a target display schematic diagram;

[0012] Analyze the display structure, extract and select features, and obtain at least one first feature, wherein the at least one first feature constitutes a first feature set;

[0013] Calling the classifier to identify the first feature set, determining whether the first feature set is in a product production completion state, and if so, performing appearance comparison, and if not, updating the first feature set according to the updated scene structure at a preset interval, and re-identifying;

[0014] The appearance information of the first feature set is obtained, and the appearance standard database is called to determine whether the size information of the first feature set is qualified. If so, no processing is performed. If not, an early warning is issued and production adjustments are made to the workstation where the first feature set is located.

[0015] Preferably, the construction of indicators for target features to obtain a shape standard database comprises the following steps:

[0016] Select the corresponding algorithm formula according to the entered size features, shape features and color features;

[0017] Calculate the appearance specification parameters corresponding to the entered size features, shape features and color features;

[0018] Generate a three-dimensional model based on the data points obtained by the algorithm formula and parameter calculation;

[0019] A shape standard database is generated based on the input size features, shape features, color features and three-dimensional models.

[0020] Preferably, the classifier design comprises the following steps:

[0021] Obtain sample monitoring images under different production processes, use the gradient formula to calculate the gradient change rate of adjacent pixel points, obtain the pixel points whose gradient change rate is greater than a first preset value as the edge of the production sample feature, and fit to generate the edge fitting function of the production sample feature;

[0022] Acquire at least one appearance sample feature of products in different production states, classify the at least one appearance sample feature according to a finished state and an unfinished state, and obtain a finished state sample set and an unfinished state sample set;

[0023] In the unfinished state sample set, the appearance sample features are classified. The classification method is to take the first production sample feature and the second production sample feature, calculate the overlapping area of ​​the first production sample feature and the second production sample feature according to the first edge fitting function and the second edge fitting function of the first production sample feature and the second production sample feature, and calculate the characteristic proportion of the overlapping area to the first production sample feature;

[0024] If the feature proportion is greater than the second preset value, the first production sample feature and the second production sample feature are classified into the same category; if the feature proportion does not exceed the second preset value, the first production sample feature and the second production sample feature are classified into different categories;

[0025] In the completed state sample set, the production sample features are classified, and the classification method is consistent with the classification method of the unfinished state sample set;

[0026] Only one production sample feature is retained among the same type of production sample features, and the remaining production sample features are deleted.

[0027] Preferably, the gradient formula is as follows:

[0028]

[0029] Among them, a is the gradient change rate, b and c are the pixel values ​​of adjacent pixels, e and g as well as f and h are the coordinates of adjacent pixels respectively.

[0030] Preferably, the classifier evaluation and repair comprises the following steps:

[0031] Obtain at least one production test feature, and determine whether the state of the production test feature is a completed state or an uncompleted state;

[0032] Obtain at least one production test feature, and determine whether the state of the production test feature is a completed state or an uncompleted state;

[0033] The production test feature is input into the classifier, and the classifier searches for the third production sample feature with the highest fit to the production test feature, and obtains the classification of the third production sample feature;

[0034] The classifier outputs a judgment result according to the third production sample feature, and the output result is a completed state or an unfinished state;

[0035] Compare the output result with the actual state of the production test feature to see if they are consistent. If yes, do nothing. If no, count once in the counter.

[0036] Count the number of counts in the counter and the number of production test features, and calculate the classification error rate;

[0037] If the classification error rate is greater than the third preset value, the classifier is repaired; if the classification error rate does not exceed the third preset value, no processing is performed.

[0038] Preferably, the stretching and transforming of the target object image to obtain the target object display schematic diagram comprises the following steps:

[0039] Get the distance from the drone camera to the target object, and calculate the zoom ratio of the target object image based on the distance;

[0040] The target object image is scaled according to the scaling ratio to obtain a target object display schematic diagram, and the size ratio of the target object display schematic diagram to the actual production site is one to one.

[0041] Preferably, the extracting and selecting features to obtain at least one first feature comprises the following steps:

[0042] Obtain a target object display schematic diagram, and perform display structure analysis according to the target object display schematic diagram;

[0043] The gradient formula is used to calculate the gradient change rate of adjacent pixel points, and the pixel points whose gradient change rate is greater than a first preset value are obtained as the edge of the first feature, and the edge fitting function of the first feature is generated by fitting.

[0044] Preferably, the determining whether the first feature set is in a product production completion state comprises the following steps:

[0045] Taking the first feature in the first feature set, the first feature is input into the classifier, and the classifier calculates the overlapping area of ​​the first feature and the shape sample feature according to the edge fitting function of the first feature and the edge fitting function of the shape sample feature, and calculates the feature ratio of the overlapping area to the first feature;

[0046] In the classifier, the shape sample feature with the largest feature proportion is selected as the fourth shape sample feature, and the state of the fourth shape sample feature is determined. If the fourth shape sample feature is in a completed state, the first feature is qualified; if the fourth shape sample feature is in an unfinished state, the first feature is unqualified;

[0047] When the first feature traverses the first feature set, if all the first features are qualified, the first feature set is in a completed state; if there is one first feature that is unqualified, the first feature set is in an unfinished state.

[0048] Preferably, calling the standard size database to determine whether the size information of the first feature set is qualified comprises the following steps:

[0049] Calling the first feature in the first feature set, obtaining the edge fitting function of the first feature, identifying the first feature as one of a size feature, a shape feature, and a color feature, and recording it as a preprocessing category;

[0050] Calculate the size of the first feature according to the edge fitting function, and calculate the absolute value of the difference between the size of the first feature and a standard value of the preprocessing category;

[0051] If the absolute value of the difference is greater than the fourth preset value, the size of the first feature is unqualified; if the absolute value of the difference does not exceed the fourth preset value, the size of the first feature is qualified;

[0052] When the first feature traverses the first feature set, if all the first features are qualified, the size information of the first feature set is judged to be qualified; if there is one first feature that is unqualified, the size information of the first feature set is judged to be unqualified.

[0053] A UAV target detection system based on deep learning is used to implement the UAV target detection method based on deep learning as described above, comprising:

[0054] A target feature extraction module, wherein the target feature extraction module obtains the size feature, shape feature and color feature of the target, constructs indicators based on the target features, and obtains a shape standard database;

[0055] A pattern recognition training module, wherein the pattern recognition training module performs classifier design and classifier evaluation and repair;

[0056] An image acquisition module, which acquires image information of a target object by photographing it through a drone, and converts the image information of the target object into electrical information;

[0057] An image preprocessing module, wherein the image preprocessing module acquires electrical information and performs image preprocessing, the preprocessing including signal enhancement, signal denoising, signal restoration and image normalization processing;

[0058] An image correction module, wherein the image correction module corrects the preprocessed target image and performs a stretching transformation on the target image to obtain a target display schematic diagram;

[0059] A display feature extraction module, wherein the scene feature extraction module analyzes the display structure, extracts and selects features, and obtains at least one first feature, wherein the at least one first feature constitutes a first feature set;

[0060] A pattern matching module, wherein the pattern matching module calls a classifier to identify a first feature set, determines whether the first feature set is in a product production completion state, and if so, performs an appearance comparison; if not, updates the first feature set according to an updated scene structure at a preset interval, and performs identification again;

[0061] The detection and comparison module obtains the appearance information of the first feature set, calls the appearance standard database, and determines whether the size information of the first feature set is qualified. If so, no processing is performed. If not, an early warning is issued and production adjustments are made to the workstation where the first feature set is located.

[0062] Compared with the prior art, the present invention has the following beneficial effects:

[0063] By setting up a pattern recognition training module, an image correction module, a display feature extraction module and a pattern matching module, the finished and unfinished states are distinguished during detection. When the product is unfinished, no appearance comparison is made. When the product is completed, an appearance comparison is made. This allows timely detection of whether the appearance of the product during production meets the requirements. For products that do not meet the requirements, production adjustments can be made as soon as possible. In addition, the image correction module stretches and transforms the target object image to obtain a target object display schematic diagram. The ratio of the target object display schematic diagram to the actual production site is one-to-one, which is convenient for judging the size from the image. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0065] Figure 2 A schematic diagram of the method flow for constructing indicators for target features and obtaining a standard appearance database in the present invention;

[0066] Figure 3 A schematic diagram of the method flow for designing a classifier in the present invention;

[0067] Figure 4 It is a flowchart of a method for determining whether a first feature set is in a product production completion state in the present invention;

[0068] Figure 5 It is a flowchart of the method for judging whether the size information of the first feature set is qualified in the present invention. DETAILED DESCRIPTION

[0069] 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.

[0070] Reference Figure 1-5 As shown, a UAV target detection method based on deep learning includes:

[0071] Obtain the size, shape and color features of the target, construct indicators based on the target features, and obtain a standard appearance database;

[0072] Design, evaluate and repair classifiers;

[0073] Acquire the target image information by shooting with a drone and convert the target image information into electrical information;

[0074] Acquire electrical information and perform image preprocessing, including signal enhancement, signal denoising, signal recovery and image normalization;

[0075] Correct the preprocessed target image, stretch and transform the target image, and obtain a target display schematic diagram;

[0076] Analyze the display structure, extract and select features, and obtain at least one first feature, wherein the at least one first feature constitutes a first feature set;

[0077] Calling the classifier to identify the first feature set, determining whether the first feature set is in a product production completion state, and if so, performing appearance comparison, and if not, updating the first feature set according to the updated scene structure at a preset interval, and re-identifying;

[0078] The appearance information of the first feature set is obtained, and the appearance standard database is called to determine whether the size information of the first feature set is qualified. If so, no processing is performed. If not, an early warning is issued and production adjustments are made to the workstation where the first feature set is located.

[0079] The construction of indicators for target features to obtain a shape standard database includes the following steps:

[0080] Select the corresponding algorithm formula according to the entered size features, shape features and color features;

[0081] Calculate the appearance specification parameters corresponding to the entered size features, shape features and color features;

[0082] Generate a three-dimensional model based on the data points obtained by the algorithm formula and parameter calculation;

[0083] Generate a shape standard database based on the input size features, shape features, color features and three-dimensional models;

[0084] The appearance standard database is used as a comparison during monitoring. When laying transmission lines, there are specific requirements for the product's size, shape and color characteristics. The accuracy during production needs to be within the allowable error range, otherwise, production failures may occur.

[0085] The classifier design comprises the following steps:

[0086] Obtain sample monitoring images under different production processes, use the gradient formula to calculate the gradient change rate of adjacent pixel points, obtain the pixel points whose gradient change rate is greater than a first preset value as the edge of the production sample feature, and fit to generate the edge fitting function of the production sample feature;

[0087] Acquire at least one appearance sample feature of products in different production states, classify the at least one appearance sample feature according to a finished state and an unfinished state, and obtain a finished state sample set and an unfinished state sample set;

[0088] In the unfinished state sample set, the appearance sample features are classified. The classification method is to take the first production sample feature and the second production sample feature, calculate the overlapping area of ​​the first production sample feature and the second production sample feature according to the first edge fitting function and the second edge fitting function of the first production sample feature and the second production sample feature, and calculate the characteristic proportion of the overlapping area to the first production sample feature;

[0089] If the feature proportion is greater than the second preset value, the first production sample feature and the second production sample feature are classified into the same category; if the feature proportion does not exceed the second preset value, the first production sample feature and the second production sample feature are classified into different categories;

[0090] In the completed state sample set, the production sample features are classified, and the classification method is consistent with the classification method of the unfinished state sample set;

[0091] Only one production sample feature is retained among the same type of production sample features, and the remaining production sample features are deleted;

[0092] The gradient formula is as follows:

[0093]

[0094] Among them, a is the gradient change rate, b and c are the pixel values ​​of adjacent pixels, e and g as well as f and h are the coordinates of adjacent pixels respectively;

[0095] The classifier is mainly used to judge the production status in the monitored image, judging whether it is an unfinished state or a completed state. The principle is to extract the appearance sample features of the unfinished state, extract the appearance sample features of the completed state, classify the appearance sample features, remove the duplication of the appearance sample features in the same category, and reduce the amount of calculation during comparison. When comparing, the closest appearance sample features are found. According to the existing unfinished state or completed state of the appearance sample features, it can be concluded that the actual production has been completed or unfinished.

[0096] The classifier evaluation and repair comprises the following steps:

[0097] Obtain at least one production test feature, and determine whether the state of the production test feature is a completed state or an uncompleted state;

[0098] Obtain at least one production test feature, and determine whether the state of the production test feature is a completed state or an uncompleted state;

[0099] The production test feature is input into the classifier, and the classifier searches for the third production sample feature with the highest fit to the production test feature, and obtains the classification of the third production sample feature;

[0100] The classifier outputs a judgment result according to the third production sample feature, and the output result is a completed state or an unfinished state;

[0101] Compare the output result with the actual state of the production test feature to see if they are consistent. If yes, do nothing. If no, count once in the counter.

[0102] Count the number of counts in the counter and the number of production test features, and calculate the classification error rate;

[0103] If the classification error rate is greater than the third preset value, the classifier is repaired; if the classification error rate does not exceed the third preset value, no processing is performed;

[0104] The purpose of classifier evaluation and repair is to enhance the fault tolerance of the classifier, because the directly generated classifier may have certain problems. Using a problematic classifier for judgment may easily lead to wrong judgments. Therefore, the classifier needs to be tested and repaired according to the test results.

[0105] Stretching and transforming the target object image to obtain a target object display schematic diagram includes the following steps:

[0106] Get the distance from the drone camera to the target object, and calculate the zoom ratio of the target object image based on the distance;

[0107] The target object image is scaled according to the scaling ratio to obtain a target object display schematic diagram, wherein the size ratio of the target object display schematic diagram to the actual production site is one to one;

[0108] Since the image is taken at a long distance, its size is different from the actual size, which will cause trouble in image processing. Therefore, the target object image is stretched and transformed to obtain a one-to-one size ratio between the target object display schematic and the actual production site. When processing the display schematic, there is no need to consider the difference between the display schematic and the actual size.

[0109] Performing feature extraction and selection to obtain at least one first feature includes the following steps:

[0110] Obtain a target object display schematic diagram, and perform display structure analysis according to the target object display schematic diagram;

[0111] The gradient formula is used to calculate the gradient change rate of adjacent pixel points, and the pixel points whose gradient change rate is greater than a first preset value are obtained as the edge of the first feature, and the edge fitting function of the first feature is generated by fitting.

[0112] Determining whether the first feature set is in a product production completion state includes the following steps:

[0113] Taking the first feature in the first feature set, the first feature is input into the classifier, and the classifier calculates the overlapping area of ​​the first feature and the shape sample feature according to the edge fitting function of the first feature and the edge fitting function of the shape sample feature, and calculates the feature ratio of the overlapping area to the first feature;

[0114] In the classifier, the shape sample feature with the largest feature proportion is selected as the fourth shape sample feature, and the state of the fourth shape sample feature is determined. If the fourth shape sample feature is in a completed state, the first feature is qualified; if the fourth shape sample feature is in an unfinished state, the first feature is unqualified;

[0115] When the first feature traverses the first feature set, if all the first features are qualified, the first feature set is in a completed state; if there is a first feature that is unqualified, the first feature set is in an unfinished state;

[0116] The first feature set has multiple first features. Using multiple first features for judgment can ensure that the judgment result has high fault tolerance and avoid misjudgment caused by a single factor.

[0117] Calling the standard size database to determine whether the size information of the first feature set is qualified includes the following steps:

[0118] Calling the first feature in the first feature set, obtaining the edge fitting function of the first feature, identifying the first feature as one of a size feature, a shape feature, and a color feature, and recording it as a preprocessing category;

[0119] Calculate the size of the first feature according to the edge fitting function, and calculate the absolute value of the difference between the size of the first feature and a standard value of the preprocessing category;

[0120] If the absolute value of the difference is greater than the fourth preset value, the size of the first feature is unqualified; if the absolute value of the difference does not exceed the fourth preset value, the size of the first feature is qualified;

[0121] When the first feature traverses the first feature set, if all the first features are qualified, the size information of the first feature set is judged to be qualified; if there is one first feature that is unqualified, the size information of the first feature set is judged to be unqualified.

[0122] A UAV target detection system based on deep learning, comprising:

[0123] A target feature extraction module, wherein the target feature extraction module obtains the size feature, shape feature and color feature of the target, constructs indicators based on the target features, and obtains a shape standard database;

[0124] A pattern recognition training module, wherein the pattern recognition training module performs classifier design and classifier evaluation and repair;

[0125] An image acquisition module, which acquires image information of a target object by photographing it through a drone, and converts the image information of the target object into electrical information;

[0126] An image preprocessing module, wherein the image preprocessing module acquires electrical information and performs image preprocessing, the preprocessing including signal enhancement, signal denoising, signal restoration and image normalization processing;

[0127] An image correction module, wherein the image correction module corrects the preprocessed target image and performs a stretching transformation on the target image to obtain a target display schematic diagram;

[0128] A display feature extraction module, wherein the scene feature extraction module analyzes the display structure, extracts and selects features, and obtains at least one first feature, wherein the at least one first feature constitutes a first feature set;

[0129] A pattern matching module, wherein the pattern matching module calls a classifier to identify a first feature set, determines whether the first feature set is in a product production completion state, and if so, performs an appearance comparison; if not, updates the first feature set according to an updated scene structure at a preset interval, and performs identification again;

[0130] The detection and comparison module obtains the appearance information of the first feature set, calls the appearance standard database, and determines whether the size information of the first feature set is qualified. If so, no processing is performed. If not, an early warning is issued and production adjustments are made to the workstation where the first feature set is located.

[0131] In summary, the advantages of the present invention are: by setting a pattern recognition training module, an image correction module, a display feature extraction module and a pattern matching module, the completed and unfinished states are distinguished during detection. When the product is unfinished, no appearance comparison is made. When the product is completed, an appearance comparison is made, so that it is possible to promptly discover whether the appearance of the product during production meets the requirements. For products that do not meet the requirements, production adjustments can be made at the first time. In addition, the image correction module stretches and transforms the target object image to obtain a target object display schematic diagram. The ratio of the target object display schematic diagram to the actual production site is one to one, which is convenient for judging the size from the image.

[0132] 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 UAV target detection method based on deep learning, characterized in that: include: Obtain the size, shape and color features of the target, construct indicators based on the target features, and obtain a standard appearance database; Design, evaluate and repair classifiers; Acquire the target image information by shooting with a drone and convert the target image information into electrical information; Acquire electrical information and perform image preprocessing, including signal enhancement, signal denoising, signal recovery and image normalization; Correct the preprocessed target image, stretch and transform the target image, and obtain a target display schematic diagram; Analyze the display structure, extract and select features, and obtain at least one first feature, wherein the at least one first feature constitutes a first feature set; Calling the classifier to identify the first feature set, determining whether the first feature set is in a product production completion state, and if so, performing appearance comparison, and if not, updating the first feature set according to the updated scene structure at a preset interval, and re-identifying; The appearance information of the first feature set is obtained, and the appearance standard database is called to determine whether the size information of the first feature set is qualified. If so, no processing is performed. If not, an early warning is issued and production adjustments are made to the workstation where the first feature set is located.

2. The method for detecting drone targets based on deep learning according to claim 1, characterized in that: The construction of indicators for target features to obtain a shape standard database includes the following steps: Select the corresponding algorithm formula according to the entered size features, shape features and color features; Calculate the appearance specification parameters corresponding to the entered size features, shape features and color features; Generate a three-dimensional model based on the data points obtained by the algorithm formula and parameter calculation; A shape standard database is generated based on the input size features, shape features, color features and three-dimensional models.

3. The method for detecting drone targets based on deep learning according to claim 2, characterized in that: The classifier design comprises the following steps: Obtain sample monitoring images under different production processes, use the gradient formula to calculate the gradient change rate of adjacent pixel points, obtain the pixel points whose gradient change rate is greater than a first preset value as the edge of the production sample feature, and fit to generate the edge fitting function of the production sample feature; Acquire at least one appearance sample feature of products in different production states, classify the at least one appearance sample feature according to a finished state and an unfinished state, and obtain a finished state sample set and an unfinished state sample set; In the unfinished state sample set, the appearance sample features are classified. The classification method is to take the first production sample feature and the second production sample feature, calculate the overlapping area of ​​the first production sample feature and the second production sample feature according to the first edge fitting function and the second edge fitting function of the first production sample feature and the second production sample feature, and calculate the characteristic proportion of the overlapping area to the first production sample feature; If the feature proportion is greater than the second preset value, the first production sample feature and the second production sample feature are classified into the same category; if the feature proportion does not exceed the second preset value, the first production sample feature and the second production sample feature are classified into different categories; In the completed state sample set, the production sample features are classified, and the classification method is consistent with the classification method of the unfinished state sample set; Only one production sample feature is retained among the same type of production sample features, and the remaining production sample features are deleted.

4. The method for detecting drone targets based on deep learning according to claim 3, characterized in that: The gradient formula is as follows: Among them, a is the gradient change rate, b and c are the pixel values ​​of adjacent pixels, e and g as well as f and h are the coordinates of adjacent pixels respectively.

5. The method for detecting drone targets based on deep learning according to claim 4, characterized in that: The classifier evaluation and repair comprises the following steps: Obtain at least one production test feature, and determine whether the state of the production test feature is a completed state or an uncompleted state; The production test feature is input into the classifier, and the classifier searches for the third production sample feature with the highest fit to the production test feature, and obtains the classification of the third production sample feature; The classifier outputs a judgment result according to the third production sample feature, and the output result is a completed state or an unfinished state; Compare the output result with the actual state of the production test feature to see if they are consistent. If yes, do nothing. If no, count once in the counter. Count the number of counts in the counter and the number of production test features, and calculate the classification error rate; If the classification error rate is greater than the third preset value, the classifier is repaired; if the classification error rate does not exceed the third preset value, no processing is performed.

6. The method for detecting drone targets based on deep learning according to claim 5, characterized in that: The stretching transformation of the target object image to obtain the target object display schematic diagram comprises the following steps: Get the distance from the drone camera to the target object, and calculate the zoom ratio of the target object image based on the distance; The target object image is scaled according to the scaling ratio to obtain a target object display schematic diagram, and the size ratio of the target object display schematic diagram to the actual production site is one to one.

7. The method for detecting drone targets based on deep learning according to claim 6, characterized in that: The feature extraction and selection to obtain at least one first feature comprises the following steps: Obtain a target object display schematic diagram, and perform display structure analysis according to the target object display schematic diagram; The gradient formula is used to calculate the gradient change rate of adjacent pixel points, and the pixel points whose gradient change rate is greater than a first preset value are obtained as the edge of the first feature, and the edge fitting function of the first feature is generated by fitting.

8. The method for detecting drone targets based on deep learning according to claim 7, characterized in that: The step of determining whether the first feature set is in a product production completion state comprises the following steps: Taking the first feature in the first feature set, the first feature is input into the classifier, and the classifier calculates the overlapping area of ​​the first feature and the shape sample feature according to the edge fitting function of the first feature and the edge fitting function of the shape sample feature, and calculates the feature ratio of the overlapping area to the first feature; In the classifier, the shape sample feature with the largest feature proportion is selected as the fourth shape sample feature, and the state of the fourth shape sample feature is determined. If the fourth shape sample feature is in a completed state, the first feature is qualified; if the fourth shape sample feature is in an unfinished state, the first feature is unqualified; When the first feature traverses the first feature set, if all the first features are qualified, the first feature set is in a completed state; if there is one first feature that is unqualified, the first feature set is in an unfinished state.

9. The method for detecting drone targets based on deep learning according to claim 8, characterized in that: The calling of the standard size database to determine whether the size information of the first feature set is qualified comprises the following steps: Calling the first feature in the first feature set, obtaining the edge fitting function of the first feature, identifying the first feature as one of a size feature, a shape feature, and a color feature, and recording it as a preprocessing category; Calculate the size of the first feature according to the edge fitting function, and calculate the absolute value of the difference between the size of the first feature and a standard value of the preprocessing category; If the absolute value of the difference is greater than the fourth preset value, the size of the first feature is unqualified; if the absolute value of the difference does not exceed the fourth preset value, the size of the first feature is qualified; When the first feature traverses the first feature set, if all the first features are qualified, the size information of the first feature set is judged to be qualified; if there is one first feature that is unqualified, the size information of the first feature set is judged to be unqualified.

10. A UAV target detection system based on deep learning, used to implement a UAV target detection method based on deep learning as described in claims 1-9, characterized in that: include: A target feature extraction module, wherein the target feature extraction module obtains the size feature, shape feature and color feature of the target, constructs indicators based on the target features, and obtains a shape standard database; A pattern recognition training module, wherein the pattern recognition training module performs classifier design and classifier evaluation and repair; An image acquisition module, which acquires image information of a target object by photographing it through a drone, and converts the image information of the target object into electrical information; An image preprocessing module, wherein the image preprocessing module acquires electrical information and performs image preprocessing, the preprocessing including signal enhancement, signal denoising, signal restoration and image normalization processing; An image correction module, wherein the image correction module corrects the preprocessed target image and performs a stretching transformation on the target image to obtain a target display schematic diagram; A display feature extraction module, wherein the scene feature extraction module analyzes the display structure, extracts and selects features, and obtains at least one first feature, wherein the at least one first feature constitutes a first feature set; A pattern matching module, wherein the pattern matching module calls a classifier to identify a first feature set, determines whether the first feature set is in a product production completion state, and if so, performs an appearance comparison; if not, updates the first feature set according to an updated scene structure at a preset interval, and performs identification again; The detection and comparison module obtains the appearance information of the first feature set, calls the appearance standard database, and determines whether the size information of the first feature set is qualified. If so, no processing is performed. If not, an early warning is issued and production adjustments are made to the workstation where the first feature set is located.

Images