A deep learning-based industrial robot visual image recognition method and system

Abstract

The present application relates to a kind of industrial robot vision image recognition method and system based on deep learning, 1, by constructing bright value, dark value two light field, different paint under different light field variation can be generated training model, so that different paint color under different light field can be presented with the change of color;2, by the information of industrial product in advance input can build input variable, eliminate the error of presented color caused by industrial product parameter;3, use the same paint in dry and wet state under the different presented color of construction model, by image recognition completes the judgment of paint dry;4, build dry judgment model and color gamut deviation model with self-learning ability complete the judgment of entire industrial product, can be based on color range different according to the degree of change of paint color before and after drying to complete the judgment of industrial product dry state, while it can avoid the situation that judgment cannot be carried out due to too small color change of paint on industrial product.

Description

Technical Field

[0001] This invention relates to industrial image recognition, and more specifically, to a method and system for visual image recognition of industrial robots based on deep learning. Background Technology

[0002] With the upgrading of industrial automation technology, industrial robots, with their high adaptability and flexibility, are widely used in intelligent automated production. Among these applications, vision technology for industrial robots has been extensively developed and applied. Generally, vision technology is used in assembly line processes such as positioning and quality inspection, but its application is less common in industrial product drying. Industrial product drying lines typically achieve their drying purpose through long drying times. The different positions and orientations of industrial products within the drying line, resulting in varying wind directions, lead to different drying efficiencies. Furthermore, the type of material being dried and the coating material also affect the drying time. Due to the potential for multiple factors... The use of shared drying lines and customized coating processes for industrial products can lead to uncertainty about drying completion time. As a result, industrial robots cannot immediately clamp the dried products. This is typically addressed by adding more redundancy to the drying line, such as increasing the dwell time of the products. However, this reduces the overall efficiency of the line. Image recognition faces significant challenges in industrial drying technology due to the large volume of color data and the varying drying values ​​of different colors on different materials and in different locations. Furthermore, each industrial product typically has multiple colors, making image recognition extremely difficult. Summary of the Invention

[0003] In view of this, the first objective of the present invention is to provide a deep learning-based method for visual image recognition of industrial robots;

[0004] The second objective of this invention is to provide a deep learning-based visual image recognition system for industrial robots.

[0005] To address the aforementioned technical problems, the technical solution of this invention is: a deep learning-based method for visual image recognition in industrial robots, comprising:

[0006] Steps for obtaining basic information: Obtaining basic information about industrial products;

[0007] The initial image acquisition step involves acquiring an initial brightness image of the industrial product under a preset brightness value light field, and an initial darkness value image under a preset darkness value light field.

[0008] In the deviation calculation step, real-time bright value images under the bright value light field and real-time dark value images under the dark value light field of the industrial product are collected at a first preset time interval. The real-time bright value images are compared with the initial bright value images to obtain the bright value deviation value of each image area. The real-time dark value images are compared with the initial dark value images to obtain the dark value deviation value of each image area. When the sum of the bright value deviation value and the dark value deviation value is higher than the first deviation benchmark, the color change judgment step is entered.

[0009] The color change judgment step calculates the consistency deviation value of each image area. The light value deviation value, dark value deviation value and consistency deviation value of each image area are brought into the color gamut deviation model through the preset color gamut deviation strategy to calculate the color change deviation of each image area. When the color change deviation of any image area is greater than the second deviation benchmark, the drying judgment step is entered; otherwise, the deviation calculation step is returned.

[0010] The drying judgment step involves inputting the color change deviation of each area into the drying judgment model to calculate the drying confidence value of the industrial product. If the drying confidence value is higher than the drying confidence threshold, the drying verification step is initiated; otherwise, the process returns to the deviation calculation step.

[0011] The drying verification step involves obtaining the actual drying results of industrial products using drying testing equipment to generate test result information.

[0012] In the model correction step, if the detection result information indicates that drying is successful, the drying judgment model and color gamut deviation model are corrected through a preset positive feedback learning strategy. If the detection result information indicates that drying has failed, the drying judgment model and color gamut deviation model are corrected through a negative feedback learning strategy.

[0013] Furthermore: the color gamut deviation model includes a color gamut reference index constructed based on industrial product basic information. Different industrial basic information corresponds to different color gamut index datasets. The color gamut index dataset includes several basic color value ranges and the light value reference deviation and dark value reference deviation corresponding to each basic color value range. The color gamut deviation strategy is to determine the corresponding color gamut index dataset from the color gamut deviation model based on the industrial product basic information input in the basic information acquisition step, and calculate the average color value of each map area through the initial light value image. Based on the basic color value range into which the average color value of the corresponding map area falls, the corresponding light value reference difference and dark value reference difference are retrieved.

[0014] The color gamut deviation model is also equipped with a color gamut deviation formula, which is used to calculate the color deviation of each color gamut.

[0015] Furthermore: A feature point localization algorithm is configured, and the deviation calculation step determines the initial values ​​of the light value and the initial values ​​of the dark value based on the image point coordinates of the image area determined by the feature point localization algorithm;

[0016] The feature point localization algorithm includes

[0017] Step A1: Determine the outline shape of the map area and determine the coordinates of the center point based on the outline shape;

[0018] Step A2: Perform depth recognition on the map area and calculate the depth value of each pixel;

[0019] Step A3: Calculate the positioning weight of each pixel according to the positioning weight formula, and select the pixel with the largest positioning weight as the coordinates of the image point in the map area. The positioning weight formula is as follows:

[0020] ,in The localization weights for the image point with coordinates (x, y) are... These are preset positioning adjustment parameters, where S is the area of ​​the map region. Let be the area of ​​the i-th segmented region in the graph. Let be the maximum curvature of the i-th segmented region in the graph. Let be the depth value of the image point with coordinates (x, y). This represents the average depth value of the map area. The coordinates of the center point are given.

[0021] Furthermore: the color gamut deviation formula includes ,in For the color deviation of the nth map region, This represents the consistency deviation value for the nth map region. The deviation calculation function is used when the deviation value is a variable. Let the deviation calculation function be the dark value deviation value as a variable, then we have , ,in This is the deviation value of the bright value. As the benchmark value, This is the dark value deviation value. is the baseline value for the dark value, and b is a preset adjustment parameter with a value range between 0.02 and 0.5.

[0022] Furthermore: the basic information of the industrial product includes a surface material item, a coating method item, and a coating material item. The surface material item reflects the surface material of the industrial product, the coating method item reflects the coating method of the industrial product, and the coating material item reflects the corresponding coating material of the industrial product.

[0023] The color gamut deviation model is configured with a preset adhesion humidity table. The adhesion humidity table is configured with adhesion value and humidity value for each surface material item, coating method item, and coating material item. The color gamut reference index of the basic information of industrial products with the same adhesion value and humidity value is the same.

[0024] Furthermore: the drying judgment model includes a color value association neural network, each node of the color value association neural network corresponds to a basic color value range, and a trust transfer value reflecting the drying correlation of color values ​​is formed between the nodes. The drying judgment model can output the corresponding trust transfer value according to the basic color value range of any two map areas.

[0025] In the drying judgment step, the trust sub-value for each map region is calculated using a drying trust formula. The drying trust value is the sum of the trust sub-values ​​for each map region. The drying trust formula is as follows:

[0026] ,in For the trust sub-value of the nth region, The trust transfer value is the value corresponding to the i-th adjacent region in the graph. Let i be the area of ​​the i-th map region adjacent to this map region. This represents the area of ​​the map region. Let be the length of the boundary line between the i-th map region and its adjacent map regions. Let be the perimeter of the area shown in the figure. The color deviation is the m-th map region adjacent to this map region.

[0027] Further: The positive feedback learning strategy includes configuring a positive trust threshold. When the difference between the dry trust value and the dry trust threshold is less than the positive trust threshold, a corresponding new trust transfer value is calculated using a first positive feedback formula. The first positive feedback formula is: ,in Passing values ​​for trust, The preset trust balancing parameters, This represents the number of nodes between two regions in a color value association neural network. The trust sub-value of the m-th map region adjacent to this map region. To ensure the dryness of the handicrafts, This is the dryness trust threshold; when the difference between the dryness trust value and the dryness trust threshold is greater than the positive trust threshold, if... The new dark value baseline is then... If The new benchmark value is ,in These are the preset color gamut approximation parameters.

[0028] Further: The negative feedback learning strategy includes configuring a color-changing reference value, determining map areas where the color-changing deviation is less than the color-changing reference value, and calculating a new trust transfer value for the corresponding map area using a first negative feedback formula, wherein the first negative feedback formula is: ,in These are the preset trust intervention parameters; if... The new benchmark value is If The new dark value baseline is then... ,in These are the preset color gamut intervention parameters.

[0029] To achieve the second objective of this invention, a deep learning-based industrial robot visual image recognition system is provided, which applies the aforementioned deep learning-based industrial robot visual image recognition method. The system is characterized by further comprising a light field construction device, an image acquisition device, and a light field feedback device. The light field construction device is used to construct a bright light field and a dark light field. The light field feedback device is coupled to the light field construction device to detect the actual illumination intensity and adjust the output of the light field construction device. The image acquisition device includes several image acquisition units positioned at different locations on the production line and an image controller. The image acquisition units acquire scene images and form corresponding images using the image stitching algorithm of the image controller.

[0030] Furthermore, the drying detection equipment includes an image acquisition unit disposed on the clamping part of the industrial robot. The image acquisition unit is used to acquire image information of the clamping part of the industrial robot and to determine the actual drying result based on image comparison.

[0031] The main technical effects of this invention are reflected in the following aspects: 1. By constructing two light fields, one for brightness and one for darkness, a training model can be generated to address the changes in different coatings under different light fields, allowing different coating colors to be presented as color changes under different light fields; 2. Input variables can be constructed using pre-input industrial product information, eliminating errors in color presentation caused by industrial product parameters; 3. By constructing a model based on the different colors of the same coating in the dry and wet states, the dryness of the coating is determined through image recognition; 4. A self-learning dryness judgment model and a color gamut deviation model are constructed to determine the dryness of the entire industrial product. The dryness state of the industrial product can be determined based on the degree of color change before and after drying, while avoiding situations where the color change of the coating on the industrial product is too small to make a judgment. Attached Figure Description

[0032] Figure 1 : Schematic diagram of the invention.

[0033] The attached figures are labeled as follows: S1, Basic Information Acquisition Step; S2, Initial Image Acquisition Step; S3, Deviation Calculation Step; S4, Color Change Judgment Step; S5, Drying Judgment Step; S6, Drying Verification Step; S7, Model Correction Step. Detailed Implementation

[0034] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings, so that the technical solution of the present invention can be more easily understood and mastered.

[0035] This application is first applied in drying systems or air-drying production lines, which are equipped with industrial robots. Cameras can be mounted on the robots or independently configured for image feedback and recognition. The purpose of this method is to determine whether industrial products have been sufficiently dried, ensuring that the mechanical gripper does not damage the paint on the product surface during handling. It also allows for targeted judgments for each product, ensuring high efficiency of the entire production line and enabling the line to simultaneously handle multiple types and shapes of industrial products. The specific design is as follows: A deep learning-based industrial robot visual image recognition method, including...

[0036] Step S1, acquiring basic information, involves obtaining basic information about the industrial product. This basic information relates to the actual color and color changes after drying. Specifically, it includes surface material, coating method, and paint material. The surface material reflects the surface material, the coating method reflects the coating method, and the paint material reflects the corresponding paint material. The user inputs pre-defined options for surface material, coating method, and paint material, and accesses a pre-set adhesion humidity table in the database. This table, configured with adhesion and humidity values ​​for each surface material, coating method, and paint material, provides the same color gamut reference index for products with the same adhesion and humidity values. The data of these three types is formatted to obtain the adhesion and humidity values. Different materials correspond to different values. By looking up the table, products with the same actual color effect can be linked to the same formatted index, thus establishing a basic data model for color changes.

[0037] In the initial image acquisition step S2, initial images of the industrial products under a preset brightness value light field and initial images of the industrial products under a preset darkness value light field are acquired respectively. The brightness value light field and the darkness value light field can be constructed by a combination of lighting and sensors. Preferably, if the image acquisition camera is directly mounted on the robotic arm, then the lighting is also mounted on the robotic arm, and the sensor is mounted on the corresponding production line. In this way, the lighting generates changes in the light field, providing a brighter lighting environment and a darker lighting environment that is close to natural light, avoiding the influence of different color rendering caused by different natural light. The darkness value light field is the first type of light field with a light intensity slightly higher than natural light, and the brightness value light field is the second type of light field with a light intensity much higher than natural light. By constructing the two light fields, errors caused by the light field due to the color itself or the material itself can be eliminated. The sensor is used to adjust the output intensity of the lighting to form the corresponding light field. Then, the camera can capture the image of each craft in the initial state, and then the image features are mapped to the basic information corresponding to the craft, while obtaining information such as the shape and color of the industrial products in the image.

[0038] The deviation calculation step S3 mainly consists of two steps. The first step is to collect real-time bright value images and real-time dark value images of industrial products under bright value light field and dark value light field at first preset time intervals, so as to monitor the changes of the images in real time and provide a basis for judgment.

[0039] The second step is to calculate the color difference. The real-time brightness image is compared with the initial brightness image to obtain the brightness deviation value of each area. The real-time darkness image is compared with the initial darkness image to obtain the darkness deviation value of each area. When the sum of the brightness deviation value and the darkness deviation value is higher than the first deviation benchmark, the color change judgment step is entered. Calculating the color difference requires extracting the color values ​​in the image. Since there are multiple colors on industrial products, the image must first be divided into corresponding areas according to the similarity of the color values ​​of each pixel in the image. This means that each area is the same color. However, in actual image acquisition, even the same area may have pixels with different color values. Therefore, preferably, one pixel in the area should represent the color value of the entire area. A feature point positioning algorithm is configured. The deviation calculation step S3 determines the initial brightness value and the initial darkness value by using the image point coordinates of the area determined by the feature point positioning algorithm.

[0040] The feature point localization algorithm includes

[0041] Step A1: Determine the outline shape of the map area and determine the coordinates of the center point based on the outline shape;

[0042] Step A2: Perform depth recognition on the map area and calculate the depth value of each pixel;

[0043] Step A3: Calculate the positioning weight of each pixel according to the positioning weight formula, and select the pixel with the largest positioning weight as the coordinates of the image point in the map area. The positioning weight formula is as follows:

[0044] ,in The localization weights for the image point with coordinates (x, y) are... The preset positioning adjustment parameters are given, where S is the area of ​​the map region. Let be the area of ​​the i-th segmented region in the graph. Let be the maximum curvature of the i-th segmented region in the graph. Let be the depth value of the image point with coordinates (x, y). This represents the average depth value of the map area. The coordinates of the center point are used. Determining the coordinates of this center point requires considering two factors. The first is the position of the center point. Theoretically, the closer this center point is to the geometric center of the area, the better it represents the color of that area. A depth model is constructed for each industrial product. Here, depth should be understood as relative depth, i.e., depth relative to the airflow direction of the drying equipment, or depth relative to the clamping direction of the clamping equipment. Depth values ​​can be obtained in two ways: one is by using multiple cameras to capture images and correlate their relative positions to form a 3D model; the second is by associating the 3D model loaded onto the industrial product itself with the feature recognition of the actual image to complete the depth recognition of the center point in the image, thereby obtaining the depth value of each coordinate. Simultaneously... The formula can calculate the positioning weight of each coordinate, and then select the coordinate with the highest weight as the coordinate point for color collection in this map area. In this way, the difference calculation can be quickly completed by the color difference of the same coordinate at different times. Then, by setting a deviation benchmark value, all color changes are summed. When the change exceeds the first deviation benchmark, the next step is taken. The purpose of this is that the subsequent loading of the model requires a large amount of computation, and if the calculation is repeated, it will consume too much computing power. Therefore, a benchmark is set. When the sum of the differences of all coordinates of the industrial product at two times is greater than the preset standard, it means that the color of the industrial product has changed at least once. The first deviation benchmark is set according to the surface area of ​​the industrial product and the number of map areas on it.

[0045] If, after a certain period of time, the surface color of the industrial product changes significantly according to the data, then the process proceeds to color change judgment step S4. The purpose of this step is to calculate the actual color change deviation of each image area based on the different changes caused by different colors and different basic information of the industrial product, using a color gamut deviation model. Specifically, this involves calculating the consistency deviation value of each image area, and using a preset color gamut deviation strategy to input the brightness deviation value, darkness deviation value, and consistency deviation value of each image area into the color gamut deviation model to calculate the color change deviation of each image area. The color gamut deviation model includes a color gamut base constructed based on the basic information of the industrial product. The quasi-index, different industrial basic information corresponds to different color gamut index datasets, the color gamut index dataset includes several basic color value ranges and the light value reference deviation and dark value reference deviation corresponding to each basic color value range; the color gamut deviation strategy is to determine the corresponding color gamut index dataset from the color gamut deviation model based on the industrial product basic information input in the basic information acquisition step, and calculate the average color value of each map area through the initial light value image, and retrieve the corresponding light value reference difference and dark value reference difference according to the basic color value range in which the average color value of the corresponding map area falls; the color gamut deviation model is also configured with a color gamut deviation formula, which is used to calculate the color change deviation of each color gamut. By importing initial information about industrial products, a corresponding index dataset is obtained. Then, the initial color of each image region is obtained by using the average color value of the initial image under a bright light field. This allows us to retrieve the required changes in bright and dark values ​​for this color after drying. The initial model is constructed by training on samples of different industrial products and colors, and by combining mathematical methods for fitting. This allows us to obtain the required changes in bright and dark values ​​for each color under different indices. Finally, the color gamut deviation formula is used... ,in For the color deviation of the nth map region, This represents the consistency deviation value for the nth map region. The deviation calculation function is used when the deviation value is a variable. Let the deviation calculation function be the dark value deviation value as a variable, then we have , ,in This is the deviation value of the bright value. As the benchmark value, This is the dark value deviation value. Here, b is the baseline value for the dark value, and 'b' is a preset adjustment parameter ranging from 0.02 to 0.5. Since the actual light and dark light fields only influence each other, the larger data point is used as the basis, with the smaller data point serving as an auxiliary variable. Simultaneously, a consistency deviation value is calculated, which is the difference between the average color value of the area and the color value at the center coordinate. This yields the color change deviation for each area. Note that the color change deviation may be negative. For example, if the actual test requires a color change of 4 units, but no color change is observed during the actual sampling, the color change deviation is -4. This situation may be due to uneven drying or an incorrect baseline deviation setting for the color material due to a small sample size. In this case, the drying judgment model needs to make a judgment. Similarly, to conserve the computational power of the drying judgment model, when the color change deviation of any area is greater than the second deviation baseline, the drying judgment step is initiated; otherwise, the deviation calculation step is returned. When the color change deviation of any area is large, it indicates that at least one area has met the color change standard, and the next step can proceed.

[0046] In the drying judgment step S5, the color change deviation of each map area is input into the drying judgment model to calculate the drying trust value of the industrial product. When the drying trust value is higher than the drying trust threshold, the drying verification step is initiated; otherwise, the process returns to the deviation calculation step. The purpose of the drying judgment step is to calculate the drying trust value of the entire industrial product. The drying judgment model includes a color value association neural network. Each node of the color value association neural network corresponds to a basic color value range. Trust transfer values ​​reflecting the drying correlation of color values ​​are formed between the nodes. The drying judgment model can output the corresponding trust transfer value based on the basic color value ranges of any two map areas.

[0047] In the drying judgment step, the trust sub-value for each map region is calculated using a drying trust formula. The drying trust value is the sum of the trust sub-values ​​for each map region. The drying trust formula is as follows:

[0048] ,in For the trust sub-value of the nth region, The trust transfer value is the value corresponding to the i-th adjacent region in the graph. Let i be the area of ​​the i-th map region adjacent to this map region. This represents the area of ​​the map region. Let be the length of the boundary line between the i-th map region and its adjacent map regions. Let be the perimeter of the area shown in the figure. This represents the color change deviation of the m-th adjacent map area. Since the trust values ​​between map areas can influence each other besides color change (meaning each map area has adjacent areas), and if one map area dries well, the adjacent areas should also dry well within the same timeframe, a trust value propagation function is constructed to eliminate errors caused by some areas not changing color. This function considers factors such as area, the length of the boundary line between map areas, and perimeter. Each map area is affected by the color change deviation of its adjacent areas, thus obtaining its own trust value. This is the significance of the drying judgment model setting. When the sum of the trust values, i.e., the drying trust value, is greater than the drying trust threshold, it can be determined that drying is complete. The drying trust threshold varies depending on the number of map areas and their surface area.

[0049] Drying verification step S6 involves obtaining the actual drying results of the industrial product using a drying testing device to generate testing result information.

[0050] In model correction step S7, if the detection result indicates successful drying, the drying judgment model and color gamut deviation model are corrected using a preset positive feedback learning strategy. If the detection result indicates drying failure, the drying judgment model and color gamut deviation model are corrected using a negative feedback learning strategy. The positive feedback learning strategy includes a configured positive trust threshold. When the difference between the drying trust value and the drying trust threshold is less than the positive trust threshold, a new trust transfer value is calculated using a first positive feedback formula. The first positive feedback formula is... ,in The preset trust balancing parameters, This represents the number of nodes between two regions in a color value association neural network. The trust sub-value of the m-th map region adjacent to this map region. To ensure the dryness confidence level of this craft, This is the dryness trust threshold; when the difference between the dryness trust value and the dryness trust threshold is greater than the positive trust threshold, if... The new dark value baseline is then... If The new benchmark value is ,in The preset color gamut approximation parameters are used. The negative feedback learning strategy includes configuring a color change reference value, identifying image areas where the color change deviation is less than the color change reference value, and calculating a new trust transfer value for the corresponding image area using a first negative feedback formula. The first negative feedback formula is... ,in These are preset trust intervention parameters; if... The new benchmark value is If The new dark value baseline is then... ,in These are preset color gamut intervention parameters. Through positive and negative feedback, this invention possesses deep learning capabilities, thus improving reliability.

[0051] In another embodiment, a deep learning-based industrial robot visual image recognition system applies the aforementioned deep learning-based industrial robot visual image recognition method. The system is characterized by further including a light field construction device, an image acquisition device, and a light field feedback device. The light field construction device is used to construct a bright light field and a dark light field. The light field feedback device is coupled to the light field construction device to detect the actual illumination intensity and adjust the output of the light field construction device. The image acquisition device includes several image acquisition units positioned at different locations on the production line and an image controller. The image acquisition units acquire scene images and form corresponding images using the image stitching algorithm of the image controller.

[0052] The drying detection equipment includes an image acquisition unit installed on the gripping part of an industrial robot. This unit acquires image information from the gripping part and determines the actual drying result based on image comparison. In other words, the image is placed on the gripping device. If drying is incomplete, paint residue will remain on the gripping device. Image comparison can identify this residue, indicating incomplete drying. While manual inspection or other detection methods can also be used, all currently high-precision detection methods are destructive. Therefore, these methods cannot directly replace the image recognition detection logic of this invention; they can only be used for training samples and feedback.

[0053] Of course, the above are just typical examples of the present invention. In addition, the present invention may have many other specific embodiments. All technical solutions formed by equivalent substitution or equivalent transformation fall within the scope of protection claimed by the present invention.

Claims

1. A deep learning-based visual image recognition method for industrial robots, characterized in that: include Steps for obtaining basic information: Obtaining basic information about industrial products; The initial image acquisition step involves acquiring an initial brightness image of the industrial product under a preset brightness value light field, and an initial darkness value image under a preset darkness value light field. In the deviation calculation step, real-time bright value images under the bright value light field and real-time dark value images under the dark value light field of the industrial product are collected at a first preset time interval. The real-time bright value images are compared with the initial bright value images to obtain the bright value deviation value of each image area. The real-time dark value images are compared with the initial dark value images to obtain the dark value deviation value of each image area. When the sum of the bright value deviation value and the dark value deviation value is higher than the first deviation benchmark, the color change judgment step is entered. The color change judgment step calculates the consistency deviation value of each image area. The light value deviation value, dark value deviation value and consistency deviation value of each image area are brought into the color gamut deviation model through the preset color gamut deviation strategy to calculate the color change deviation of each image area. When the color change deviation of any image area is greater than the second deviation benchmark, the drying judgment step is entered; otherwise, the deviation calculation step is returned. The drying judgment step involves inputting the color change deviation of each area into the drying judgment model to calculate the drying confidence value of the industrial product. If the drying confidence value is higher than the drying confidence threshold, the drying verification step is initiated; otherwise, the process returns to the deviation calculation step. The drying verification step involves obtaining the actual drying results of industrial products using drying testing equipment to generate test result information. In the model correction step, if the detection result information indicates that drying is successful, the drying judgment model and color gamut deviation model are corrected through a preset positive feedback learning strategy. If the detection result information indicates that drying has failed, the drying judgment model and color gamut deviation model are corrected through a negative feedback learning strategy. 2.The industrial robot vision image recognition method based on deep learning of claim 1, wherein: The color gamut deviation model includes a color gamut reference index constructed based on industrial product basic information. Different industrial basic information corresponds to different color gamut index datasets. The color gamut index dataset includes several basic color value ranges and the light value reference deviation and dark value reference deviation corresponding to each basic color value range. The color gamut deviation strategy is to determine the corresponding color gamut index dataset from the color gamut deviation model based on the industrial product basic information input in the basic information acquisition step, and calculate the average color value of each map area through the initial light value image. Based on the basic color value range in which the average color value of the corresponding map area falls, the corresponding light value reference difference and dark value reference difference are retrieved. The color gamut deviation model is also equipped with a color gamut deviation formula, which is used to calculate the color deviation of each color gamut. 3.The industrial robot vision image recognition method based on deep learning of claim 2, wherein: The system is equipped with a feature point localization algorithm, and the deviation calculation step determines the initial values ​​of the bright and dark values ​​by using the image point coordinates of the image area determined by the feature point localization algorithm. The feature point localization algorithm includes Step A1: Determine the outline shape of the map area and determine the coordinates of the center point based on the outline shape; Step A2: Perform depth recognition on the map area and calculate the depth value of each pixel; Step A3: Calculate the positioning weight of each pixel according to the positioning weight formula, and select the pixel with the largest positioning weight as the coordinates of the image point in the map area. The positioning weight formula is as follows: ,in The localization weights for the image point with coordinates (x, y) are... These are preset positioning adjustment parameters, where S is the area of ​​the map region. Let be the area of ​​the i-th segmented region in the graph. Let be the maximum curvature of the i-th segmented region in the graph. Let be the depth value of the image point with coordinates (x, y). This represents the average depth value of the map area. The coordinates of the center point are given.

4. The deep learning-based industrial robot visual image recognition method as described in claim 3, characterized in that: The color gamut deviation formula includes ,in For the color deviation of the nth map region, This represents the consistency deviation value for the nth map region. The deviation calculation function is used when the deviation value is a variable. Let the deviation calculation function be the dark value deviation value as a variable, then we have , ,in This is the deviation value of the bright value. As the benchmark value, This is the dark value deviation value. is the baseline value for the dark value, and b is a preset adjustment parameter with a value range between 0.02 and 0.

5.

5. The deep learning-based visual image recognition method for industrial robots as described in claim 2, characterized in that: The basic information of the industrial product includes a surface material item, a coating method item, and a coating material item. The surface material item reflects the surface material of the industrial product, the coating method item reflects the coating method of the industrial product, and the coating material item reflects the corresponding coating material of the industrial product. The color gamut deviation model is configured with a preset adhesion humidity table. The adhesion humidity table is configured with adhesion value and humidity value for each surface material item, coating method item, and coating material item. The color gamut reference index of the basic information of industrial products with the same adhesion value and humidity value is the same.

6. The deep learning-based visual image recognition method for industrial robots as described in claim 1, characterized in that: The drying judgment model includes a color value association neural network. Each node of the color value association neural network corresponds to a basic color value range. A trust transfer value reflecting the drying correlation of color values ​​is formed between the nodes. The drying judgment model can output the corresponding trust transfer value according to the basic color value range of any two map areas. In the drying judgment step, the trust sub-value for each map region is calculated using a drying trust formula. The drying trust value is the sum of the trust sub-values ​​for each map region. The drying trust formula is as follows: ,in For the trust sub-value of the nth region, The trust transfer value is the value corresponding to the i-th adjacent region in the graph. Let i be the area of ​​the i-th map region adjacent to this map region. This represents the area of ​​the map region. Let be the length of the boundary line between the i-th map region and its adjacent map regions. Let be the perimeter of the area shown in the figure. The color deviation is the m-th map region adjacent to this map region.

7. The deep learning-based visual image recognition method for industrial robots as described in claim 3, characterized in that: The positive feedback learning strategy includes configuring a positive trust threshold. When the difference between the dry trust value and the dry trust threshold is less than the positive trust threshold, a corresponding new trust transfer value is calculated using a first positive feedback formula. The first positive feedback formula is: ,in Passing values ​​for trust, The preset trust balancing parameters, This represents the number of nodes between two regions in a color value association neural network. The trust sub-value of the m-th map region adjacent to this map region. To ensure the dryness confidence value of this industrial product, This is the dryness trust threshold; when the difference between the dryness trust value and the dryness trust threshold is greater than the positive trust threshold, if... The new dark value baseline is then... If The new benchmark value is ,in These are preset color gamut approximation parameters.

8. The deep learning-based visual image recognition method for industrial robots as described in claim 7, characterized in that: The negative feedback learning strategy includes configuring a color-changing benchmark value, identifying map areas where the color-changing deviation is less than the benchmark value, and calculating a new trust transfer value for the corresponding map areas using a first negative feedback formula. The first negative feedback formula is... ,in These are preset trust intervention parameters; if... The new benchmark value is If The new dark value baseline is then... ,in These are the preset color gamut intervention parameters.

9. A deep learning-based industrial robot visual image recognition system, employing the deep learning-based industrial robot visual image recognition method according to any one of claims 1-8, characterized in that: It also includes a light field construction device, an image acquisition device, and a light field feedback device. The light field construction device is used to construct a bright light field and a dark light field. The light field feedback device is coupled to the light field construction device to detect the actual light intensity and adjust the output of the light field construction device. The image acquisition device includes several image acquisition units set at different positions on the production line and an image controller. The image acquisition units acquire scene images and form corresponding images through the image stitching algorithm of the image controller.

10. The deep learning-based industrial robot visual image recognition system as described in claim 9, characterized in that: The drying detection equipment includes an image acquisition unit installed in the clamping part of the industrial robot. The image acquisition unit is used to acquire image information of the clamping part of the industrial robot and to determine the actual drying result based on image comparison.

Images