Industrial robot visual positioning method and system based on machine learning

Abstract

The invention relates to the technical field of industrial visual image recognition, in particular to an industrial robot visual positioning method and system based on machine learning, and the method comprises the steps: firstly obtaining a preset image recognition model and a target direction recognition model, and then training the target direction recognition model in combination with the preset image recognition model; and then obtaining an actual picture of the target object, obtaining an actual pose of the target object based on the trained target direction identification model, and finally performing visual positioning on the target object according to the actual pose of the target object. Compared with the prior art, the target direction recognition model is directly used for recognizing the actual pose of the target object, and the calculation amount of actual application is greatly simplified. Besides, the target direction recognition model is further trained in an auxiliary mode through the preset picture recognition model, the efficiency of the whole process from development to application is improved, and the problem that in the prior art, the use experience of the visual positioning technology for the industrial robot is not ideal enough is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial visual image recognition, and in particular to a machine learning-based industrial robot visual positioning method and system. Background Art

[0002] Industrial robots are indispensable automation equipment in modern industry. They are programmed to perform highly repetitive tasks to improve production efficiency and output, while reducing human errors and improving operational safety. Visual positioning technology is a crucial auxiliary technology for industrial robots.

[0003] Visual positioning technology is a technology that uses visual information to determine the position of an object in space. It usually involves computer vision and image processing technology, which identifies the position, direction and motion state of an object by analyzing and interpreting visual data. In the prior art, visual positioning technology generally includes multiple steps such as image recognition, feature extraction, feature comparison, and coordinate transformation.

[0004] However, in reality, the working scenes of industrial robots are relatively simple, generally guiding and correcting deviations, simple assembly, etc., and the images that need to be recognized during work are also of high similarity. Therefore, for industrial robots, the existing technology is relatively redundant. On the one hand, complex calculation steps may reduce the speed of visual recognition, thereby affecting the speed of production operations. On the other hand, the cumbersome recognition process will also increase the difficulty of optimizing the algorithm strategy for specific scenarios, making the efficiency of the actual application of the existing technology not ideal. Summary of the invention

[0005] Therefore, the present invention provides an industrial robot visual positioning method and system based on machine learning, so as to solve the problem that the visual positioning technology in the prior art has a less than ideal user experience for industrial robots.

[0006] The present invention provides an industrial robot visual positioning method based on machine learning, comprising:

[0007] Obtain a preset image recognition model and a target direction recognition model, wherein the preset image recognition model is used to recognize the probability that the image content is a target object in a preset posture, and the probability is negatively correlated with the deviation between the actual posture of the target object and the preset posture, and the target direction recognition model is a neural network model, which is used to recognize the actual posture of the target object in the image;

[0008] The target direction recognition model is trained in combination with the preset image recognition model, wherein the optimization strength of each iteration of the target direction recognition model is positively correlated with the probability output by the preset image recognition model;

[0009] Get the actual image of the target object, and obtain the actual position and posture of the target object based on the trained target direction recognition model;

[0010] Visually locate the target object according to its actual position and posture.

[0011] In a preferred implementation: the preset image recognition model is a visual word bag model, which includes multiple visual words, each visual word corresponds to a contour feature, and the preset image recognition model is used to obtain the probability that the image content is a target object in a preset position based on the number of multiple visual words contained in the image.

[0012] In a preferred implementation, training a target direction recognition model in combination with a preset image recognition model includes:

[0013] Get the preset pose vector, target training image and corresponding real label;

[0014] Preprocess the target training image to obtain the input image;

[0015] Input the input image into the preset image recognition model to obtain the object recognition probability;

[0016] Input the input image into the target direction recognition model to obtain the direction recognition vector;

[0017] Establish a loss function based on the preset pose vector, the true label, the direction recognition vector and the object recognition probability;

[0018] The target direction recognition model is optimized according to the loss function.

[0019] In a preferred implementation, preprocessing the target training image to obtain an input image includes:

[0020] Scaling the target training image to obtain a first initial image;

[0021] Offset the target training image to obtain a second initial image;

[0022] The first initial image and the second initial image are preprocessed respectively to obtain a plurality of input images.

[0023] In a preferred implementation, a loss function is established according to a preset pose vector, a true label, a direction recognition vector, and an object recognition probability, including:

[0024] According to the direction recognition vector and the true label, a first loss function is established;

[0025] According to the difference between the direction recognition vector and the preset posture vector, a difference feature value is established;

[0026] According to the numerical level difference between the difference feature value and the object recognition probability, the direction recognition vector is feature scaled to obtain a correction vector;

[0027] Establish a second loss function based on the correction vector and the true label;

[0028] A loss function is established based on the first loss function and the second loss function.

[0029] In a preferred implementation, according to the difference in numerical level between the difference feature value and the object recognition probability, feature scaling is performed on the direction recognition vector to obtain a correction vector, including:

[0030] According to the numerical level gap between the difference feature value and the object recognition probability, the scaling factor is obtained;

[0031] According to the scaling factor, the direction identification vector is corrected to increase the difference between different elements in the direction identification vector to obtain a corrected vector.

[0032] In a preferred implementation, the loss function is:

[0033] L = (1-λ) × L real +λ×L extra ;

[0034] Among them, L represents the loss function, L real represents the first loss function, L extra represents the second loss function, λ is the preset weight;

[0035] The second loss function is specifically:

[0036]

[0037] Among them, i represents the number of elements, n represents the total number of elements, and r i Represents the i-th element in the true label, y i represents the i-th element in the correction vector;

[0038] The correction vector is expressed as:

[0039]

[0040] Among them, x i represents the i-th element in the direction recognition vector, and S represents the difference eigenvalue.

[0041] In a preferred implementation, the calculation formula of the difference characteristic value is:

[0042]

[0043] Among them, x represents the direction recognition vector, r represents the true label, θ(x,r) represents the vector angle between the direction recognition vector and the true label, and P represents the object recognition probability.

[0044] In a preferred implementation, visually locating the target object according to the actual position and posture of the target object includes:

[0045] Adjust the position of the target shooting device according to the actual position and posture of the target object;

[0046] Based on the target shooting device, shoot the target object at the adjusted position to obtain a secondary collection picture;

[0047] According to the secondary collected pictures, the target object is visually located based on the preset picture recognition model.

[0048] The present invention also provides an industrial robot visual positioning system based on machine learning, comprising:

[0049] A model preparation module, used to obtain a preset image recognition model and a target direction recognition model, wherein the preset image recognition model is used to recognize the probability that the image content is a target object in a preset posture, and the probability is negatively correlated with the deviation between the actual posture of the target object and the preset posture, and the target direction recognition model is a neural network model, which is used to recognize the actual posture of the target object in the image;

[0050] A model optimization module, used to train a target direction recognition model in combination with a preset image recognition model, wherein the optimization strength of each iteration of the target direction recognition model is positively correlated with the probability output by the preset image recognition model;

[0051] The posture recognition module is used to obtain the actual image of the target object and obtain the actual posture of the target object based on the trained target direction recognition model;

[0052] The visual positioning module is used to visually locate the target object according to its actual position and posture.

[0053] The beneficial effects of adopting the above embodiment are:

[0054] The present invention provides a machine learning-based industrial robot visual positioning method and system, which first obtains a preset picture recognition model and a target direction recognition model, then trains the target direction recognition model in combination with the preset picture recognition model, then obtains an actual picture of the target object, obtains the actual position and posture of the target object based on the trained target direction recognition model, and finally visually locates the target object according to the actual position and posture of the target object. Compared with the prior art, because the application scenario of industrial robots in practice is single, the present invention directly uses the target direction recognition model to identify the actual position and posture of the target object, without the need to perform feature recognition, extraction and other steps on the picture itself, which greatly simplifies the amount of calculation in practical application. In addition, in practice, the objects to be recognized by industrial robots are often unchanged, but the viewing angles are different, which means that the probability of a preset picture recognition model recognizing an object is negatively correlated with the deviation between the actual position and the preset position. The present invention takes advantage of this point, further assists in training the target direction recognition model through the preset picture recognition model, improves the efficiency of the whole process from development to application, and solves the problem that the visual positioning technology in the prior art is not ideal for the use experience of industrial robots. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 A method flow chart of the industrial robot visual positioning method based on machine learning provided by the present invention;

[0056] Figure 2 for Figure 1 Specific step diagram of step S102;

[0057] Figure 3 for Figure 2 Specific step diagram of step S205;

[0058] Figure 4 This is a system structure diagram of the industrial robot visual positioning system based on machine learning provided by the present invention. DETAILED DESCRIPTION

[0059] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] Combination Figure 1 As shown, a specific embodiment of the present invention discloses a visual positioning method for an industrial robot based on machine learning, comprising:

[0061] S101, obtaining a preset image recognition model and a target direction recognition model, wherein the preset image recognition model is used to recognize the probability that the image content is a target object in a preset posture, and the probability is negatively correlated with the deviation between the actual posture of the target object and the preset posture, and the target direction recognition model is a neural network model, which is used to recognize the actual posture of the target object in the image;

[0062] S102, training a target direction recognition model in combination with a preset image recognition model, wherein the optimization strength of each iteration of the target direction recognition model is positively correlated with the probability output by the preset image recognition model;

[0063] S103, obtaining an actual image of the target object, and obtaining the actual position and posture of the target object based on the trained target direction recognition model;

[0064] S104: Visually locate the target object according to its actual position and posture.

[0065] Among them, the target object is an object with a shape to be identified in practice, such as a specific workpiece, a hole, a crack or edge flatness on a structure. The preset image recognition model is a preset model for identifying the target object, but it has certain requirements for the posture of the target object in the image it recognizes, that is, the preset posture. In practice, the preset image recognition model can be trained and provided by the equipment supplier. The target direction recognition model is used to identify the specific position of the target object in the image, that is, the actual posture. It can be understood that the preset posture and actual posture are relative concepts. Depending on the actual situation, it can refer to the position and posture of the target object, or it can refer to the shooting angle of the shooting device.

[0066] For industrial robots, they often need to perform a lot of simple and repetitive work, and the images they recognize often have a high similarity. In most cases, the ultimate goal of industrial robots is not to identify the specific type of the target object, but to analyze the actual position and posture of the target object in order to guide the actions of other devices. This premise provides feasibility for the neural network model to directly identify the position and posture. Therefore, this embodiment adopts the target direction recognition model to directly identify the actual position and posture of the target object to improve the operation efficiency of the algorithm during actual work. Similarly, this premise also makes the output result of the preset image recognition model more linear. In theory, when the target object is in a preset position and posture, the recognition result output by the preset image recognition model, that is, the probability that the image content is the target object is 100%, and the greater the deviation of the actual position and posture of the target object from the preset position and posture, the greater the deformation of the target object reflected in the image, and the lower the probability of the output of the preset image recognition model. The output probability is regarded as a measure of the accuracy of the output result of the target direction recognition model, which is used to assist the training of the target direction recognition model, and can greatly improve the training speed and accuracy of the target direction recognition model.

[0067] For example, if the probability of the preset image recognition model output is high, it intuitively indicates that the probability of an image including a target object is high. In the present embodiment, it can be understood that the difference between the actual position of the target object and the preset position is small. At this time, if the difference between the result of the target direction recognition input and the preset position is large, the learning rate can be increased to speed up the convergence speed. Conversely, if the difference between the result of the target direction recognition input and the preset position is small, the learning rate can be reduced to prevent overfitting.

[0068] Therefore, compared with the prior art, the present invention directly uses the target direction recognition model to identify the actual position and posture of the target object, without the need to perform feature recognition and extraction steps on the image itself, which greatly simplifies the amount of computation required for practical applications. In addition, the present invention further assists in training the target direction recognition model by presetting the image recognition model, thereby improving the efficiency of the entire process from development to application, and solving the problem that the visual positioning technology in the prior art is not ideal for the use experience of industrial robots.

[0069] Furthermore, in a preferred embodiment, the preset image recognition model is a bag-of-visual-words model, which includes a plurality of visual words, each visual word corresponding to a contour feature, and the preset image recognition model is used to obtain the probability that the image content is a target object in a preset position based on the number of the plurality of visual words contained in the image.

[0070] Bag of Visual Words (BoVW) is an image representation method widely used in the field of computer vision, which draws on the bag of words model in natural language processing. This method regards the image as a set of "visual words", similar to a set of words in a text, without considering the spatial arrangement order of these words. For this embodiment, the bag of visual words technology has a natural advantage in generating probabilities that have a linear relationship with posture deviations. For example, assuming that a variety of tiny line segments of a certain length (straight lines with different extension angles, curves with different radians, broken lines with different angles, etc.) are contour features and are used as visual words, then pictures of target objects in different postures can be represented by a variety of different visual words. By counting the distribution of these visual words, the probability that the picture content is the target object can be obtained.

[0071] Further, combined with Figure 2 As shown, in a preferred embodiment, step S102, training a target direction recognition model in combination with a preset image recognition model, specifically includes:

[0072] S201, obtaining a preset pose vector, a target training image, and a corresponding real label;

[0073] S202, preprocessing the target training image to obtain an input image;

[0074] S203, inputting the input image into a preset image recognition model to obtain an object recognition probability;

[0075] S204, inputting the input image into the target direction recognition model to obtain a direction recognition vector;

[0076] S205, establishing a loss function according to the preset posture vector, the true label, the direction recognition vector and the object recognition probability;

[0077] S206. Optimize the target direction recognition model according to the loss function.

[0078] In the above process, the target training image is the image actually used to train the target direction model, the true label is the vector representing the actual posture of the target training image, the preset posture vector is the vector representing the preset posture, and the direction recognition vector is the vector representing the posture of the target object predicted by the target direction recognition model. The specific encoding method of these vectors representing posture can be flexibly set according to actual needs. For example, two unit vectors can be used to determine the position angle of a part. At this time, the above vectors are all 6-dimensional vectors. The above process uses the preset image recognition model to assist learning by optimizing the loss function. Compared with the method of adjusting the learning rate mentioned above, its correction is more detailed and accurate.

[0079] In a preferred embodiment, the above step S202, preprocessing the target training image to obtain the input image, specifically includes:

[0080] Scaling the target training image to obtain a first initial image;

[0081] Offset the target training image to obtain a second initial image;

[0082] The first initial image and the second initial image are preprocessed respectively to obtain a plurality of input images.

[0083] For image recognition technology, when the image size is fixed and there is no pre-specified position information, the algorithm will only obtain analysis results based on the distribution of the pixels themselves. This will result in different recognition results for target images with the same posture if they appear in different positions or sizes in the image. Therefore, in the above process, in addition to the conventional equalization and binarization processes, two additional steps, scaling and offset, are added when preprocessing the target training images. The purpose is to simulate the different sizes and positions of target objects caused by changes in posture during shooting in actual application scenarios of industrial robots, and to increase the number of samples to further eliminate errors.

[0084] Further, combined with Figure 3 As shown, in a preferred embodiment, the above step S205, based on the preset posture vector, the true label, the direction recognition vector and the object recognition probability, establishes a loss function, which specifically includes:

[0085] S301, establishing a first loss function according to the direction recognition vector and the true label;

[0086] S302, establishing a difference feature value according to the difference between the direction recognition vector and the preset posture vector;

[0087] S303, performing feature scaling on the direction recognition vector according to the numerical level difference between the difference feature value and the object recognition probability to obtain a correction vector;

[0088] S304, establishing a second loss function according to the correction vector and the true label;

[0089] S305: Establish a loss function according to the first loss function and the second loss function.

[0090] In the above process, the first loss function can be regarded as the definition of the loss function in the prior art, that is, the loss obtained by analyzing the direction recognition vector itself, while the second loss function is added in the present invention on the basis of the prior art. The second loss function is established based on the difference eigenvalue, and the difference eigenvalue is used to represent the difference between the direction recognition vector and the preset posture vector, such as the vector angle between the two vectors, the cosine similarity between the two vectors, the Euclidean distance, etc. The numerical level difference refers to the difference between the direction recognition vector represented by the difference eigenvalue and the preset posture vector, and the difference between the actual posture of the target object represented by the object recognition probability and the preset posture. The difference between the two degrees of difference, that is, the difference in the difference level represented by the numerical values ​​of the two data, the difference eigenvalue and the object recognition probability, how to calculate the difference can be flexibly set according to the actual situation.

[0091] The present invention performs feature scaling on the direction recognition vector by using the numerical level difference between the difference feature value and the object recognition probability to increase or decrease the difference between different elements in the direction recognition vector, and establishes a second loss function based on this. In this way, the second loss function can be regarded as combining the probability output by the preset image recognition model to further amplify or reduce the difference between different elements in the direction recognition vector and the true label, obtain a correction vector containing more correction information, and analyze the loss obtained by the correction vector.

[0092] For example, when the difference represented by the difference eigenvalue is large, the numerical value of each element in the direction recognition vector can be adjusted to make the difference between each element in the direction recognition vector and the true label larger, so as to increase the loss size and thus increase the correction strength of the target direction recognition model, and vice versa.

[0093] The present invention provides a more preferred method. Specifically, in a preferred embodiment, the above step S303, according to the difference in the numerical level between the difference feature value and the object recognition probability, performs feature scaling on the direction recognition vector to obtain a correction vector, specifically includes:

[0094] According to the numerical level gap between the difference feature value and the object recognition probability, the scaling factor is obtained;

[0095] According to the scaling factor, the direction identification vector is corrected to increase the difference between different elements in the direction identification vector to obtain a corrected vector.

[0096] The main purpose of the above process is to increase the differences between different elements in the direction recognition vector to maintain the overall scale invariance.

[0097] In essence, the present invention is an improvement on the existing technology of knowledge distillation. In knowledge distillation, a loss function is added to use the teacher model to assist the training of the student model. This embodiment actually adds a second loss function to enable the target direction recognition model to learn the relevant knowledge of identifying the target object and the preset posture in the preset image recognition model, thereby achieving faster convergence and improving the flexibility of the application of machine learning technology.

[0098] Specifically, in a preferred embodiment, the loss function is:

[0099] L = (1-λ) × L real +λ×L extra ;

[0100] Among them, L represents the loss function, L real represents the first loss function, L extra represents the second loss function, λ is the preset weight;

[0101] The second loss function is specifically:

[0102]

[0103] Among them, i represents the number of elements, n represents the total number of elements, and r i Represents the i-th element in the true label, y i represents the i-th element in the correction vector;

[0104] The correction vector is expressed as:

[0105]

[0106] Among them, x i represents the i-th element in the direction recognition vector, and S represents the difference eigenvalue;

[0107] The calculation formula of the difference eigenvalue is:

[0108]

[0109] Among them, x represents the direction recognition vector, r represents the true label, θ(x,r) represents the vector angle between the direction recognition vector and the true label, and P represents the object recognition probability.

[0110] Further, in a preferred embodiment, the above step S104, visually positioning the target object according to the actual position and posture of the target object, specifically includes:

[0111] Adjust the position of the target shooting device according to the actual position and posture of the target object;

[0112] Based on the target shooting device, shoot the target object at the adjusted position to obtain a secondary collection picture;

[0113] According to the secondary collected pictures, the target object is visually located based on the preset picture recognition model.

[0114] It is understandable that the above process is only a preferred visual positioning method. In practice, after knowing the actual posture, other methods can be used according to specific circumstances. In this embodiment, the target shooting device is first adjusted to a position where the target object with a preset posture can be photographed, so that the existing preset image recognition model can be called for visual positioning, which improves code reusability, reduces the bloat of the actual software, and reduces deployment costs.

[0115] Combination Figure 4 As shown, the present invention also provides an industrial robot visual positioning system based on machine learning, comprising:

[0116] The model preparation module 410 is used to obtain a preset image recognition model and a target direction recognition model, wherein the preset image recognition model is used to recognize the probability that the image content is a target object in a preset posture, and the probability is negatively correlated with the deviation between the actual posture of the target object and the preset posture, and the target direction recognition model is a neural network model, which is used to recognize the actual posture of the target object in the image;

[0117] A model optimization module 420 is used to train a target direction recognition model in combination with a preset image recognition model, wherein the optimization strength of each iteration of the target direction recognition model is positively correlated with the probability output by the preset image recognition model;

[0118] A posture recognition module 430 is used to obtain an actual image of the target object and obtain the actual posture of the target object based on the trained target direction recognition model;

[0119] The visual positioning module 440 is used to perform visual positioning on the target object according to the actual position and posture of the target object.

[0120] It should be noted here that the corresponding system provided in the above embodiments can implement the technical solutions described in the above method embodiments. The specific implementation principles of the above modules or units can be found in the corresponding contents in the above method embodiments, which will not be repeated here.

[0121] The present invention provides a machine learning-based industrial robot visual positioning method and system, which first obtains a preset picture recognition model and a target direction recognition model, then trains the target direction recognition model in combination with the preset picture recognition model, then obtains an actual picture of the target object, obtains the actual position and posture of the target object based on the trained target direction recognition model, and finally visually locates the target object according to the actual position and posture of the target object. Compared with the prior art, because the application scenario of industrial robots in practice is single, the present invention directly uses the target direction recognition model to identify the actual position and posture of the target object, without the need to perform feature recognition, extraction and other steps on the picture itself, which greatly simplifies the amount of calculation in practical application. In addition, in practice, the objects to be recognized by industrial robots are often unchanged, but the viewing angles are different, which means that the probability of a preset picture recognition model recognizing an object is negatively correlated with the deviation between the actual position and the preset position. The present invention takes advantage of this point, further assists in training the target direction recognition model through the preset picture recognition model, improves the efficiency of the whole process from development to application, and solves the problem that the visual positioning technology in the prior art is not ideal for the use experience of industrial robots.

[0122] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0123] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A visual positioning method for industrial robots based on machine learning, characterized in that: include: Obtain a preset image recognition model and a target direction recognition model, wherein the preset image recognition model is used to recognize the probability that the image content is a target object in a preset posture, and the probability is negatively correlated with the deviation between the actual posture of the target object and the preset posture, and the target direction recognition model is a neural network model, which is used to recognize the actual posture of the target object in the image; The target direction recognition model is trained in combination with the preset image recognition model, wherein the optimization strength of each iteration of the target direction recognition model is positively correlated with the probability output by the preset image recognition model; Get the actual image of the target object, and obtain the actual position and posture of the target object based on the trained target direction recognition model; Visually locate the target object according to its actual position and posture.

2. The industrial robot visual positioning method based on machine learning according to claim 1 is characterized in that: The preset image recognition model is a bag-of-visual-words model. The bag-of-visual-words model includes multiple visual words, each visual word corresponds to a contour feature, and the preset image recognition model is used to obtain the probability that the image content is a target object in a preset position based on the number of multiple visual words contained in the image.

3. The industrial robot visual positioning method based on machine learning according to claim 1, characterized in that: Combine the preset image recognition model to train the target direction recognition model, including: Get the preset pose vector, target training image and corresponding real label; Preprocess the target training image to obtain the input image; Input the input image into the preset image recognition model to obtain the object recognition probability; Input the input image into the target direction recognition model to obtain the direction recognition vector; Establish a loss function based on the preset pose vector, the true label, the direction recognition vector and the object recognition probability; The target direction recognition model is optimized according to the loss function.

4. The method for industrial robot visual positioning based on machine learning according to claim 3, characterized in that: Preprocess the target training image to obtain the input image, including: Scaling the target training image to obtain a first initial image; Offset the target training image to obtain a second initial image; The first initial image and the second initial image are preprocessed respectively to obtain a plurality of input images.

5. The method for industrial robot visual positioning based on machine learning according to claim 3, characterized in that: According to the preset pose vector, true label, direction recognition vector and object recognition probability, a loss function is established, including: According to the direction recognition vector and the true label, a first loss function is established; According to the difference between the direction recognition vector and the preset posture vector, a difference feature value is established; According to the numerical level difference between the difference feature value and the object recognition probability, the direction recognition vector is feature scaled to obtain a correction vector; Establish a second loss function based on the correction vector and the true label; A loss function is established based on the first loss function and the second loss function.

6. The method for industrial robot visual positioning based on machine learning according to claim 5, characterized in that: According to the numerical level difference between the difference feature value and the object recognition probability, the direction recognition vector is feature scaled to obtain a correction vector, including: According to the numerical level gap between the difference feature value and the object recognition probability, the scaling factor is obtained; According to the scaling factor, the direction identification vector is corrected to increase the difference between different elements in the direction identification vector to obtain a corrected vector.

7. The method for industrial robot visual positioning based on machine learning according to claim 6, characterized in that: The loss function is: L=(1-λ)×L real +λ×L extra ; Among them, L represents the loss function, L real represents the first loss function, L extra represents the second loss function, λ is the preset weight; The second loss function is specifically: Among them, i represents the number of elements, n represents the total number of elements, and r i Represents the i-th element in the true label, y i represents the i-th element in the correction vector; The correction vector is expressed as: Among them, x i represents the i-th element in the direction recognition vector, and S represents the difference eigenvalue.

8. The method for industrial robot visual positioning based on machine learning according to claim 7, characterized in that: The calculation formula of the difference eigenvalue is: Among them, x represents the direction recognition vector, r represents the true label, θ(x,r) represents the vector angle between the direction recognition vector and the true label, and P represents the object recognition probability.

9. The industrial robot visual positioning method based on machine learning according to claim 1, characterized in that: Visually locate the target object based on its actual position and posture, including: Adjust the position of the target shooting device according to the actual position and posture of the target object; Based on the target shooting device, shoot the target object at the adjusted position to obtain a secondary collection picture; According to the secondary collected pictures, the target object is visually located based on the preset picture recognition model.

10. An industrial robot visual positioning system based on machine learning, characterized in that: include: A model preparation module, used to obtain a preset image recognition model and a target direction recognition model, wherein the preset image recognition model is used to recognize the probability that the image content is a target object in a preset posture, and the probability is negatively correlated with the deviation between the actual posture of the target object and the preset posture, and the target direction recognition model is a neural network model, which is used to recognize the actual posture of the target object in the image; A model optimization module, used to train a target direction recognition model in combination with a preset image recognition model, wherein the optimization strength of each iteration of the target direction recognition model is positively correlated with the probability output by the preset image recognition model; The posture recognition module is used to obtain the actual image of the target object and obtain the actual posture of the target object based on the trained target direction recognition model; The visual positioning module is used to visually locate the target object according to its actual position and posture.

Images