Glass surface stain image recognition method applied to window cleaning robot

Abstract

A glass surface stain image recognition method applied to a window-cleaning robot relates to the technical field of image recognition, and comprises the following steps: acquiring image data, environmental perception data and tactile data of a target glass area according to a sensor terminal of the window-cleaning robot; constructing a shooting parameter lookup table, and updating shooting parameters of the sensor terminal in real time according to the environment sensing data in the current acquisition period and the shooting parameter lookup table; constructing a stain feature recognition model, and performing pre-estimated stain type labeling on the image data of the current acquisition period according to the stain feature recognition model; and performing correctness detection on the estimated stain type marked in the target glass area according to the tactile data of the current acquisition period, if a wrong stain type exists in the target glass area, obtaining the recognition correct rate of the stain feature recognition model, and performing local feature enhanced recognition operation or retraining operation according to the recognition correct rate. And the stain identification capability of the window cleaning robot is obviously improved.

Description

Technical Field

[0001] The invention relates to the technical field of image recognition, and in particular to a glass surface stain image recognition method applied to a window cleaning robot. Background Art

[0002] CN116758051A "Glass surface defect detection method and equipment based on artificial intelligence image recognition" includes the following steps: step S1, obtain quality objection defect pictures and perform preprocessing; step S2, use the preprocessed pictures as input of the defect category determination model to obtain a first determination result on the defect type, and the defect category determination model is a classification model; step S3, use the first determination result and the preprocessed pictures as input of the defect severity determination model to obtain a second determination result on the defect severity; step S4, store the original quality objection defect pictures and the first determination result and the second determination result together in the quality objection database, and output the results.

[0003] CN115631190A "Glass substrate defect identification and warning method, system, electronic device and storage medium" collects images of the target liquid crystal glass substrate when the display background is a plurality of different colors as the initial image; pre-processes the multiple images for each color corresponding to the image to be detected to obtain the image to be detected corresponding to the color; identifies the type of surface defects of the target liquid crystal glass substrate according to the images to be detected corresponding to various colors; determines the warning mode according to the type of surface defects of the target liquid crystal glass substrate to issue a warning.

[0004] Existing window cleaning robots mainly rely on fixed cleaning paths and simple stain detection methods, which are difficult to adapt to different environments and complex stain situations. Traditional image recognition methods perform poorly in the face of problems such as lighting changes, reflections, and background interference, resulting in low accuracy and robustness in stain recognition. Therefore, a more efficient and accurate stain image recognition method is needed to improve the intelligence level and cleaning effect of window cleaning robots. Summary of the invention

[0005] In order to solve the above technical problems, the object of the present invention is to provide a glass surface stain image recognition method applied to a window cleaning robot, comprising the following steps:

[0006] Step s1: According to the sensor terminal of the window cleaning robot, the image data, environmental perception data and tactile data of the target glass area are acquired, the acquisition time is marked, and the acquisition cycle is set;

[0007] Step s2: construct a shooting parameter lookup table, and update the shooting parameters of the sensor terminal in real time according to the environmental perception data and the shooting parameter lookup table in the current acquisition cycle;

[0008] Step s3: construct a stain feature recognition model, and perform an estimated stain type labeling on the image data of the current acquisition cycle according to the stain feature recognition model;

[0009] Step s4: Perform a correctness check on the estimated stain type marked in the target glass area based on the tactile data of the current acquisition cycle. If an incorrect stain type exists in the target glass area, obtain the recognition accuracy of the stain feature recognition model, and perform a local feature enhancement recognition operation or a retraining operation based on the recognition accuracy.

[0010] Furthermore, the process of constructing a shooting parameter lookup table includes:

[0011] Based on the idea of ​​simulation learning, a simulator is used to generate simulated light environments and glass sample areas corresponding to different lighting parameters, and standard image data of the glass sample areas are obtained. Then, the sensor terminal of the window-cleaning robot is simulated to shoot the glass sample areas according to different shooting parameters under different simulated light environment conditions, and the image sample data corresponding to different simulated light environment conditions and different shooting parameters are obtained. Each image sample data is compared with the standard image data to obtain the mean square error between each image sample data and the standard image data. According to the simulated light environment, shooting parameters and mean square error corresponding to each image sample data, a shooting parameter lookup table is constructed.

[0012] Furthermore, the process of updating the shooting parameters of the sensor terminal in real time according to the environmental perception data and the shooting parameter lookup table in the current collection cycle includes:

[0013] Extract the lighting parameters in the environmental perception data within the current acquisition cycle, input the lighting parameters into a shooting parameter lookup table for data retrieval, obtain the similarity between different simulated light environments and the lighting parameters in the shooting parameter lookup table, mark the simulated light environment with the highest similarity as the first simulated light environment, obtain all shooting parameters and mean square errors corresponding to the first simulated light environment, filter out the shooting parameters corresponding to the lowest mean square error from all mean square errors, and update the shooting parameters of the sensor terminal according to the shooting parameters.

[0014] Furthermore, the process of constructing a stain feature recognition model includes:

[0015] A stain feature recognition model is constructed based on deep learning, a simulator is used to generate a number of sample image data including different types of stains, the sample image data is preprocessed in data format, a gray level co-occurrence matrix of each pixel in the sample image data after data format preprocessing is obtained, and a texture feature set of the sample image data is extracted according to the gray level co-occurrence matrix;

[0016] The texture feature sets corresponding to the sample image data of different stain types are used as training sets and test sets. The training set is input into the stain feature recognition model for training until the loss function training is stable, and the model parameters are saved. The stain feature recognition model is tested with the test set until it meets the preset requirements, and the stain feature recognition model is output.

[0017] Furthermore, the process of estimating the stain type labeling of the image data of the current acquisition period according to the stain feature recognition model includes:

[0018] Performing data format preprocessing on the image data of the current acquisition cycle, obtaining a gray level co-occurrence matrix of each pixel in the image data after the data format preprocessing, extracting a texture feature set of the image data according to the gray level co-occurrence matrix, inputting the texture feature set of the image data into a stain feature recognition model, and outputting an estimated stain type corresponding to the gray level co-occurrence matrix of each pixel according to the stain feature recognition model;

[0019] The coverage of the gray level co-occurrence matrix of each pixel in the target glass area is obtained, and the coverage of the gray level co-occurrence matrix of each pixel is marked with the estimated stain type according to the estimated stain type corresponding to the gray level co-occurrence matrix of each pixel.

[0020] Furthermore, the process of checking the correctness of the estimated stain type marked in the target glass area according to the tactile data of the current acquisition cycle includes:

[0021] Constructing a stain lookup table, wherein the stain lookup table includes standard tactile data corresponding to different stain types, and obtaining standard tactile data corresponding to the estimated stain type in the target glass area according to the estimated stain type in the target glass area and the stain lookup table;

[0022] Acquire tactile data within a coverage range of an estimated stain type of a target glass area in a current acquisition cycle, perform similarity matching on the tactile data within the coverage range of the estimated stain type and standard tactile data corresponding to the estimated stain type, acquire similarity between the tactile data within the coverage range of the estimated stain type and the standard tactile data, and compare the similarity with a preset similarity threshold;

[0023] If the similarity is less than the similarity threshold, marking the estimated stain type within the coverage range to which the tactile data corresponding to the similarity belongs as an erroneous stain type;

[0024] If the similarity is greater than an equal similarity threshold, the estimated stain type within the coverage range to which the tactile data corresponding to the similarity belongs is marked as a correct stain type.

[0025] Furthermore, if there is an erroneous stain type in the target glass area, the recognition accuracy of the stain feature recognition model is obtained, and the process of performing a local feature enhancement recognition operation or a retraining operation according to the recognition accuracy includes:

[0026] If there is an incorrect stain type in the target glass area, the recognition accuracy of the stain feature recognition model is obtained according to the total coverage area of ​​the incorrect stain type and the total coverage area of ​​the correct stain type in the target glass area, and the recognition accuracy is compared with a preset recognition accuracy threshold;

[0027] If the recognition accuracy is greater than or equal to the recognition accuracy threshold, then extract the tactile data within the coverage range of the wrong stain type, perform similarity matching between the tactile data and the standard tactile data in the stain lookup table, obtain the similarity between the tactile data and each standard tactile data in the stain lookup table, screen out the corresponding standard tactile data with a similarity greater than the similarity threshold, obtain the stain type corresponding to the standard tactile data, mark the stain type as a key stain type, and perform a local feature enhancement recognition operation according to the key stain type;

[0028] If the recognition accuracy is less than the recognition accuracy threshold, the texture feature set within the coverage range of the wrong stain type is extracted, and the stain type is annotated for the coverage range of the wrong stain type by manual annotation, a feature database is constructed, and the texture feature set within the coverage range of the wrong stain type and the stain type annotated by manual annotation are associated and stored in the feature database;

[0029] The stain types in the feature database and the texture feature sets associated with the stain types are added to the training set and the test set of the stain feature recognition model. The stain feature recognition model is retrained according to the training set and the test set, and the trained stain feature recognition model is output.

[0030] Furthermore, the process of performing local feature enhancement recognition operation according to the key stain type includes:

[0031] Constructing a local feature recognition enhancement model, obtaining model parameters of the stain feature recognition model, and using the model parameters of the stain feature recognition model as initial model parameters of the local feature recognition enhancement model;

[0032] Using a simulator to generate a plurality of sample image data including different types of stains, each of which includes a key type of stain, performing data format preprocessing on the sample image data to obtain a texture feature set of the sample image data;

[0033] The texture feature set corresponding to each sample image data is used as a training set and a test set. The training set is input into the local feature recognition enhancement model for training until the loss function training is stable, and the model parameters are saved. The local feature recognition enhancement model is tested by the test set until it meets the preset requirements, and the local feature recognition enhancement model is output.

[0034] The glass surface stain image recognition method applied to the window cleaning robot proposed in the present invention aims to significantly improve the stain recognition ability and cleaning efficiency of the window cleaning robot through a series of technical innovations and optimizations. The beneficial effects of the present invention are described in detail below:

[0035] 1. Enhance the accuracy and robustness of stain recognition. By constructing a shooting parameter lookup table, the present invention can dynamically adjust the shooting parameters of the sensor terminal according to the lighting conditions of the current environment, ensuring that clear and accurate image data can be obtained under different lighting conditions. This adaptive adjustment mechanism greatly improves the accuracy and robustness of stain recognition, especially in complex and changeable outdoor environments.

[0036] 2. Improve the generalization ability of the stain feature recognition model. The present invention adopts deep learning technology to construct the stain feature recognition model, and trains it through a large amount of simulated sample image data. These sample image data cover various stain types and texture features, ensuring that the model can accurately identify different types of stains in practical applications. In addition, the texture features are extracted through the gray-level co-occurrence matrix, which further enhances the model's recognition ability for subtle texture differences.

[0037] 3. Real-time and high efficiency. The present invention realizes real-time recognition of stains on the glass surface by real-time updating of shooting parameters and dynamic adjustment of recognition models. This method not only improves the recognition speed, but also ensures the timeliness and accuracy of the recognition results. Especially in large-scale cleaning tasks, this real-time performance can significantly improve the work efficiency of the window cleaning robot.

[0038] 4. Multimodal data fusion improves recognition accuracy. The present invention not only relies on image data, but also combines environmental perception data and tactile data. Through multimodal data fusion, the accuracy of stain recognition is further improved. The introduction of tactile data enables the model to confirm the type of stain through tactile feedback when visual information is insufficient, greatly reducing the probability of misrecognition.

[0039] 5. Adaptive learning and continuous optimization. The present invention designs local feature enhancement recognition operations and **retraining operations. When the model recognizes the wrong type of stain, it can automatically perform local feature enhancement or retraining. This adaptive learning mechanism ensures that the model can continuously optimize and improve its recognition ability when facing new types of stains or complex environments. In addition, through manual annotation and the construction of feature databases, the training data is further enriched, and the generalization ability and adaptability of the model are improved.

[0040] In summary, the present invention significantly improves the stain recognition ability of the window cleaning robot through a series of technical innovations and optimizations, and has important practical value and broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a schematic diagram of a method for recognizing stains on a glass surface applied to a window cleaning robot according to an embodiment of the present application. DETAILED DESCRIPTION

[0042] The following is a clear and complete description of the technical solutions in the embodiments of the present application in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present application.

[0043] like Figure 1 As shown, the glass surface stain image recognition method applied to the window cleaning robot includes the following steps:

[0044] Step s1: according to the sensor terminal of the window cleaning robot, image data, environmental perception data and tactile data of the target glass area are obtained, the acquisition time is marked, and the acquisition cycle is set. The sensor terminal includes an ambient light sensor, a visual sensor and a tactile sensor. The ambient light sensor is used to monitor the surrounding light conditions in real time. The visual sensor is used to obtain image data and environmental perception data of the target glass area. The tactile sensor is used to sense the humidity, hardness and viscosity of the target glass area.

[0045] Step s2: construct a shooting parameter lookup table, and update the shooting parameters of the sensor terminal in real time according to the environmental perception data and the shooting parameter lookup table in the current acquisition cycle;

[0046] Step s3: construct a stain feature recognition model, and perform an estimated stain type labeling on the image data of the current acquisition cycle according to the stain feature recognition model;

[0047] Step s4: Perform a correctness check on the estimated stain type marked in the target glass area based on the tactile data of the current acquisition cycle. If an incorrect stain type exists in the target glass area, obtain the recognition accuracy of the stain feature recognition model, and perform a local feature enhancement recognition operation or a retraining operation based on the recognition accuracy.

[0048] It should be further explained that, in the specific implementation process, the process of constructing the shooting parameter lookup table includes:

[0049] Based on the idea of ​​simulation learning, a simulator is used to generate simulated light environments and glass sample areas corresponding to different lighting parameters (including illumination, light quality, etc.), and standard image data of the glass sample area is obtained. Then, the sensor terminal of the window cleaning robot is simulated to shoot the glass sample area under different simulated light environment conditions according to different shooting parameters (including exposure time, white balance, ISO sensitivity, etc.), and the image sample data corresponding to different simulated light environment conditions and different shooting parameters are obtained. Each image sample data is compared with the standard image data to obtain the mean square error between each image sample data and the standard image data. According to the simulated light environment, shooting parameters and mean square error corresponding to each image sample data, a shooting parameter lookup table is constructed.

[0050] It should be further explained that, in the specific implementation process, the specific process of obtaining the mean square error between each image sample data and the standard image data is:

[0051] Read the image sample data I need to compare a and standard image data I b , and they have the same size M×N, the image sample data I a and standard image data I b Each pixel value in is marked as I a (x, y) and I b (x, y), the formula for calculating the mean square error is:

[0052]

[0053] Here, MES stands for mean square error.

[0054] It should be further explained that, in the specific implementation process, the process of updating the shooting parameters of the sensor terminal in real time according to the environmental perception data and the shooting parameter lookup table in the current collection cycle includes:

[0055] Extract the lighting parameters in the environmental perception data within the current acquisition cycle, input the lighting parameters into a shooting parameter lookup table for data retrieval, obtain the similarity between different simulated light environments and the lighting parameters in the shooting parameter lookup table, mark the simulated light environment with the highest similarity as the first simulated light environment, obtain all shooting parameters and mean square errors corresponding to the first simulated light environment, filter out the shooting parameters corresponding to the lowest mean square error from all mean square errors, and update the shooting parameters of the sensor terminal according to the shooting parameters.

[0056] It should be further explained that, in the specific implementation process, the process of building a stain feature recognition model includes:

[0057] A stain feature recognition model is constructed based on deep learning, a simulator is used to generate a number of sample image data including different types of stains, the sample image data is preprocessed in data format, a gray level co-occurrence matrix of each pixel in the sample image data after data format preprocessing is obtained, and a texture feature set of the sample image data is extracted according to the gray level co-occurrence matrix;

[0058] It should be noted that the image data captured by the sensor terminal may have noise problems. Therefore, in one embodiment of the present invention, after the image data is obtained, it is necessary to perform image preprocessing operations on it to improve the quality of the image data so as to obtain image data for subsequent analysis. Improving the image quality through image preprocessing operations is an image processing method well known to those skilled in the art. In the embodiment of the invention, the noise in the image is removed by a Gaussian filtering method, and then the image data is equalized to improve the contrast within the image, and the image data is grayed to convert the color image into a grayscale image. The graying process is usually achieved by averaging the weights of the color channels, wherein denoising and equalization processing, semantic segmentation network and graying are technical means well known to those skilled in the art, and the specific steps are not repeated here. Other image preprocessing means may also be used in other embodiments of the present invention, which are not limited here.

[0059] The process of extracting the texture feature set of the sample image data according to the gray level co-occurrence matrix includes:

[0060] Quantize the grayscale level of the image data into discrete grayscale levels, divide the grayscale value range into several levels, for example, an 8-bit image can be divided into 16, 32, and 64 levels, and define the parameters required for the grayscale co-occurrence matrix, including distance (d) and direction (θ), which are determined according to actual application requirements. The specific steps are not repeated here. Then, based on the obtained grayscale co-occurrence matrix, extract the texture feature value corresponding to the grayscale co-occurrence matrix of each pixel point, and construct a texture feature set. The texture features in the texture feature set include but are not limited to contrast, correlation, entropy, energy, etc.;

[0061] The texture feature sets corresponding to the sample image data of different stain types (such as oil stains, dust, etc.) are used as training sets and test sets, and the training set is input into the stain feature recognition model for training until the loss function training is stable, and the model parameters are saved. The stain feature recognition model is tested with the test set until it meets the preset requirements, and the stain feature recognition model is output.

[0062] It should be further explained that, in the specific implementation process, the process of estimating the stain type labeling of the image data of the current acquisition period according to the stain feature recognition model includes:

[0063] Performing data format preprocessing on the image data of the current acquisition cycle, obtaining a gray level co-occurrence matrix of each pixel in the image data after the data format preprocessing, extracting a texture feature set of the image data according to the gray level co-occurrence matrix, inputting the texture feature set of the image data into a stain feature recognition model, and outputting an estimated stain type corresponding to the gray level co-occurrence matrix of each pixel according to the stain feature recognition model;

[0064] The coverage of the gray level co-occurrence matrix of each pixel in the target glass area is obtained, and the coverage of the gray level co-occurrence matrix of each pixel is marked with the estimated stain type according to the estimated stain type corresponding to the gray level co-occurrence matrix of each pixel.

[0065] It should be further explained that, in a specific implementation process, the process of detecting the correctness of the estimated stain type marked in the target glass area according to the tactile data of the current acquisition cycle includes:

[0066] Constructing a stain lookup table, wherein the stain lookup table includes standard tactile data corresponding to different stain types, and obtaining standard tactile data corresponding to the estimated stain type in the target glass area according to the estimated stain type in the target glass area and the stain lookup table;

[0067] Acquire tactile data within a coverage range of an estimated stain type of a target glass area in a current acquisition cycle, perform similarity matching on the tactile data within the coverage range of the estimated stain type and standard tactile data corresponding to the estimated stain type, acquire similarity between the tactile data within the coverage range of the estimated stain type and the standard tactile data, and compare the similarity with a preset similarity threshold;

[0068] If the similarity is less than the similarity threshold, marking the estimated stain type within the coverage range to which the tactile data corresponding to the similarity belongs as an erroneous stain type;

[0069] If the similarity is greater than an equal similarity threshold, the estimated stain type within the coverage range to which the tactile data corresponding to the similarity belongs is marked as a correct stain type.

[0070] It should be further explained that, in a specific implementation process, if there is an erroneous stain type in the target glass area, the process of obtaining the recognition accuracy of the stain feature recognition model and performing a local feature enhancement recognition operation or a retraining operation according to the recognition accuracy includes:

[0071] If there is an incorrect stain type in the target glass area, then the total coverage area A of the incorrect stain type in the target glass area is calculated. e and coverage of the correct stain type Total area A c , get the recognition accuracy RA of the stain feature recognition model, where, Compare the recognition accuracy with a preset recognition accuracy threshold;

[0072] If the recognition accuracy is greater than or equal to the recognition accuracy threshold, then extract the tactile data within the coverage range of the wrong stain type, perform similarity matching between the tactile data and the standard tactile data in the stain lookup table, obtain the similarity between the tactile data and each standard tactile data in the stain lookup table, screen out the corresponding standard tactile data with a similarity greater than the similarity threshold, obtain the stain type corresponding to the standard tactile data, mark the stain type as a key stain type, and perform a local feature enhancement recognition operation according to the key stain type;

[0073] If the recognition accuracy is less than the recognition accuracy threshold, the texture feature set within the coverage range of the wrong stain type is extracted, and the stain type is annotated for the coverage range of the wrong stain type by manual annotation, a feature database is constructed, and the texture feature set within the coverage range of the wrong stain type and the stain type annotated by manual annotation are associated and stored in the feature database;

[0074] The stain types in the feature database and the texture feature sets associated with the stain types are added to the training set and the test set of the stain feature recognition model. The stain feature recognition model is retrained according to the training set and the test set, and the trained stain feature recognition model is output.

[0075] It should be further explained that, in the specific implementation process, the process of performing local feature enhancement recognition operation according to the key stain type includes:

[0076] Constructing a local feature recognition enhancement model, obtaining model parameters of the stain feature recognition model, and using the model parameters of the stain feature recognition model as initial model parameters of the local feature recognition enhancement model;

[0077] Using a simulator to generate a plurality of sample image data including different types of stains, each of which includes a key type of stain, performing data format preprocessing on the sample image data to obtain a texture feature set of the sample image data;

[0078] The texture feature set corresponding to each sample image data is used as a training set and a test set. The training set is input into the local feature recognition enhancement model for training until the loss function training is stable, and the model parameters are saved. The local feature recognition enhancement model is tested by the test set until it meets the preset requirements, and the local feature recognition enhancement model is output.

[0079] The local feature recognition enhancement model is used to perform feature recognition on the target glass area, thereby significantly improving the accuracy of stain recognition on the glass surface.

[0080] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A glass surface stain image recognition method applied to a window cleaning robot, characterized in that: The following steps are involved: Step s1: According to the sensor terminal of the window cleaning robot, the image data, environmental perception data and tactile data of the target glass area are acquired, the acquisition time is marked, and the acquisition cycle is set; Step s2: construct a shooting parameter lookup table, and update the shooting parameters of the sensor terminal in real time according to the environmental perception data and the shooting parameter lookup table in the current acquisition cycle; Step s3: construct a stain feature recognition model, and perform an estimated stain type labeling on the image data of the current acquisition cycle according to the stain feature recognition model; Step s4: Perform a correctness check on the estimated stain type marked in the target glass area based on the tactile data of the current acquisition cycle. If an incorrect stain type exists in the target glass area, obtain the recognition accuracy of the stain feature recognition model, and perform a local feature enhancement recognition operation or a retraining operation based on the recognition accuracy.

2. The glass surface stain image recognition method applied to a window cleaning robot according to claim 1, characterized in that: The process of building a shooting parameter lookup table includes: Based on the idea of ​​simulation learning, a simulator is used to generate simulated light environments and glass sample areas corresponding to different lighting parameters, and standard image data of the glass sample areas are obtained. Then, the sensor terminal of the window-cleaning robot is simulated to shoot the glass sample areas according to different shooting parameters under different simulated light environment conditions, and the image sample data corresponding to different simulated light environment conditions and different shooting parameters are obtained. Each image sample data is compared with the standard image data to obtain the mean square error between each image sample data and the standard image data. According to the simulated light environment, shooting parameters and mean square error corresponding to each image sample data, a shooting parameter lookup table is constructed.

3. The glass surface stain image recognition method applied to a window cleaning robot according to claim 2, characterized in that: The process of updating the shooting parameters of the sensor terminal in real time according to the environmental perception data and the shooting parameter lookup table in the current collection cycle includes: Extract the lighting parameters in the environmental perception data within the current acquisition cycle, input the lighting parameters into a shooting parameter lookup table for data retrieval, obtain the similarity between different simulated light environments and the lighting parameters in the shooting parameter lookup table, mark the simulated light environment with the highest similarity as the first simulated light environment, obtain all shooting parameters and mean square errors corresponding to the first simulated light environment, filter out the shooting parameters corresponding to the lowest mean square error from all mean square errors, and update the shooting parameters of the sensor terminal according to the shooting parameters.

4. The glass surface stain image recognition method applied to a window cleaning robot according to claim 3 is characterized in that: The process of building a stain feature recognition model includes: A stain feature recognition model is constructed based on deep learning, a simulator is used to generate a number of sample image data including different types of stains, the sample image data is preprocessed in data format, a gray level co-occurrence matrix of each pixel in the sample image data after data format preprocessing is obtained, and a texture feature set of the sample image data is extracted according to the gray level co-occurrence matrix; The texture feature sets corresponding to the sample image data of different stain types are used as training sets and test sets. The training set is input into the stain feature recognition model for training until the loss function training is stable, and the model parameters are saved. The stain feature recognition model is tested with the test set until it meets the preset requirements, and the stain feature recognition model is output.

5. The glass surface stain image recognition method applied to a window cleaning robot according to claim 4, characterized in that: The process of estimating the stain type labeling of the image data of the current acquisition cycle according to the stain feature recognition model includes: Performing data format preprocessing on the image data of the current acquisition cycle, obtaining a gray level co-occurrence matrix of each pixel in the image data after the data format preprocessing, extracting a texture feature set of the image data according to the gray level co-occurrence matrix, inputting the texture feature set of the image data into a stain feature recognition model, and outputting an estimated stain type corresponding to the gray level co-occurrence matrix of each pixel according to the stain feature recognition model; The coverage of the gray level co-occurrence matrix of each pixel in the target glass area is obtained, and the coverage of the gray level co-occurrence matrix of each pixel is marked with the estimated stain type according to the estimated stain type corresponding to the gray level co-occurrence matrix of each pixel.

6. The glass surface stain image recognition method applied to a window cleaning robot according to claim 5, characterized in that: The process of checking the correctness of the estimated stain type marked in the target glass area based on the tactile data of the current acquisition cycle includes: Constructing a stain lookup table, wherein the stain lookup table includes standard tactile data corresponding to different stain types, and obtaining standard tactile data corresponding to the estimated stain type in the target glass area according to the estimated stain type in the target glass area and the stain lookup table; Acquire tactile data within a coverage range of an estimated stain type of a target glass area in a current acquisition cycle, perform similarity matching on the tactile data within the coverage range of the estimated stain type and standard tactile data corresponding to the estimated stain type, acquire similarity between the tactile data within the coverage range of the estimated stain type and the standard tactile data, and compare the similarity with a preset similarity threshold; If the similarity is less than the similarity threshold, marking the estimated stain type within the coverage range to which the tactile data corresponding to the similarity belongs as an erroneous stain type; If the similarity is greater than an equal similarity threshold, the estimated stain type within the coverage range to which the tactile data corresponding to the similarity belongs is marked as a correct stain type.

7. The glass surface stain image recognition method applied to a window cleaning robot according to claim 6, characterized in that: If there is an incorrect stain type in the target glass area, the recognition accuracy of the stain feature recognition model is obtained, and the process of performing a local feature enhancement recognition operation or a retraining operation according to the recognition accuracy includes: If there is an incorrect stain type in the target glass area, the recognition accuracy of the stain feature recognition model is obtained according to the total coverage area of ​​the incorrect stain type and the total coverage area of ​​the correct stain type in the target glass area, and the recognition accuracy is compared with a preset recognition accuracy threshold; If the recognition accuracy is greater than or equal to the recognition accuracy threshold, then extract the tactile data within the coverage range of the wrong stain type, perform similarity matching between the tactile data and the standard tactile data in the stain lookup table, obtain the similarity between the tactile data and each standard tactile data in the stain lookup table, screen out the corresponding standard tactile data with a similarity greater than the similarity threshold, obtain the stain type corresponding to the standard tactile data, mark the stain type as a key stain type, and perform a local feature enhancement recognition operation according to the key stain type; If the recognition accuracy is less than the recognition accuracy threshold, the texture feature set within the coverage range of the wrong stain type is extracted, and the stain type is annotated for the coverage range of the wrong stain type by manual annotation, a feature database is constructed, and the texture feature set within the coverage range of the wrong stain type and the stain type annotated by manual annotation are associated and stored in the feature database; The stain types in the feature database and the texture feature sets associated with the stain types are added to the training set and the test set of the stain feature recognition model. The stain feature recognition model is retrained according to the training set and the test set, and the trained stain feature recognition model is output.

8. The glass surface stain image recognition method applied to a window cleaning robot according to claim 7, characterized in that: The process of local feature enhancement recognition based on key stain types includes: Constructing a local feature recognition enhancement model, obtaining model parameters of the stain feature recognition model, and using the model parameters of the stain feature recognition model as initial model parameters of the local feature recognition enhancement model; Using a simulator to generate a plurality of sample image data including different types of stains, each of which includes a key type of stain, performing data format preprocessing on the sample image data to obtain a texture feature set of the sample image data; The texture feature set corresponding to each sample image data is used as a training set and a test set. The training set is input into the local feature recognition enhancement model for training until the loss function training is stable, and the model parameters are saved. The local feature recognition enhancement model is tested by the test set until it meets the preset requirements, and the local feature recognition enhancement model is output.

Images