A method and system for identifying coking areas on the inner wall of a milk powder drying tower

Abstract

The present application provides a kind of milk powder drying tower inner wall coking area image recognition method, by drying tower inner wall image, based on the preset three-dimensional model is carried out cylindrical coordinate reverse projection and obtains plane development diagram;Image is decomposed into specular and diffuse reflection component using bidirectional reflectance distribution function model, and the intrinsic image of diffuse reflection is retained;The image is input into encoder-decoder network preliminary segmentation, and the channel attention vector generated by deep feature is used to weight shallow feature in its skip connection, form semantic-guided texture feature map and fuse with decoder feature, output coking area probability map;Further construct conditional random field model and carry out fine segmentation, one-dimensional potential function is determined by pixel probability, binary potential function fuses label smoothness constraint and distance constraint based on intrinsic image, and the segmentation result is optimized by minimizing energy function;Independent connected region in segmentation map is post-processed, and according to the preset geometric feature threshold and the average probability value of region, false coking area is screened out, and the final recognition result is determined.

Description

Technical Field

[0001] The present application belongs to the field of image recognition, and in particular relates to a method and system for image recognition of the coking area on the inner wall of a milk powder drying tower. Background Art

[0002] A core process in milk powder production is spray drying, where atomized liquid is exposed to hot air within a high-temperature drying tower, rapidly dehydrating it to form a powdered product. During this process, due to uneven temperature distribution within the tower and the adhesion of the material to the walls, some milk powder particles can remain on the high-temperature drying tower walls for extended periods, undergoing Maillard reactions or caramelization, gradually forming a darker, harder slag. The formation and accumulation of slag can lead to a series of problems. It not only reduces the heat transfer efficiency of the drying tower and affects production stability, but also, if it falls off and mixes with the finished product, it can severely affect the sensory quality and solubility of the milk powder, become an impurity in the product, and even cause safety accidents due to excessive heat accumulation. Therefore, timely and accurate monitoring of slag on the drying tower walls is crucial for ensuring product quality, improving production efficiency, and ensuring safe production.

[0003] Currently, monitoring coke buildup on the inner walls of drying towers primarily relies on regular manual inspections. Inspection results rely heavily on worker experience, making missed detections and misidentifications prone to occur. Furthermore, frequent inspections disrupt production, reduce equipment utilization, and increase production costs. Furthermore, manual access to the towers also presents safety risks. To overcome the drawbacks of manual inspection, the industry has begun experimenting with automated monitoring using machine vision technology. However, the complex internal environment of drying towers presents significant challenges for image recognition. The stainless steel inner wall has strong specular reflectivity, creating large, high-brightness reflective areas under illumination. These highlights can obscure true coke buildup or lead to misidentification as coke. Early coke buildup has a color and texture that differs only slightly from the background, making its features less distinct. Furthermore, the coke buildup is irregular in shape and has blurred edges, making it difficult for traditional image segmentation algorithms to accurately distinguish it from interfering objects such as the background, welds, or stains. Furthermore, due to the cylindrical shape of the tower wall, directly captured two-dimensional images suffer from severe geometric distortion, hindering subsequent quantitative analysis. Therefore, how to overcome the interference of strong reflections, effectively extract the features of small targets and achieve accurate segmentation of irregular coking areas is a technical problem that needs to be solved urgently in this field. Summary of the Invention

[0004] In response to the above problems, the present application proposes a method for image recognition of the coking area on the inner wall of a milk powder drying tower, which is characterized by comprising the following steps:

[0005] Acquire an image of the inner wall of the drying tower to be identified; perform a cylindrical coordinate inverse projection transformation on the inner wall image based on a preset three-dimensional geometric model of the inner wall of the drying tower to obtain a planar unfolded image; utilize a physics-based bidirectional reflectance distribution function (BRDF) model to decompose the planar unfolded image into a specular reflection component and a diffuse reflection component, and eliminate the specular reflection component to obtain a diffuse reflection component map as an intrinsic image;

[0006] The intrinsic image is input into an encoder-decoder network for preliminary segmentation, wherein, in the skip connection path of the network, the channel attention vector generated by the deep feature map of the encoder is used to weight the shallow feature map of the encoder to generate a semantically guided texture feature map, and the semantically guided texture feature map is fused with the feature map of the corresponding layer of the decoder, and the network outputs a preliminary focus area probability map;

[0007] A conditional random field model is constructed to perform fine segmentation on the probability map, wherein a unary potential function of the conditional random field model is determined by the probability that each pixel in the probability map belongs to the focus area, and a binary potential function combines the label smoothness constraint between adjacent pixels and the inter-pixel distance constraint calculated based on the intrinsic image. A refined focus area segmentation map is obtained by minimizing the energy function of the conditional random field model;

[0008] The independent connected regions in the refined coking region segmentation map are post-processed, and the independent connected regions are screened according to the preset geometric feature thresholds and regional average probability values ​​of each connected region, and pseudo coking regions that do not conform to the physical characteristics of coking are eliminated to obtain the recognition results of the coking regions.

[0009] Preferably, the step of decomposing the planar expanded image into a specular reflection component and a diffuse reflection component using a physics-based bidirectional reflectance distribution function (BRDF) model specifically includes: using a two-color reflection model to decompose the pixel value I of any pixel point in the planar expanded image into a linear superposition of a diffuse reflection component M and a specular reflection component S; obtaining an initial diffuse reflection component map by calculating the minimum color channel value of the image and performing a morphological closing operation on the channel image; estimating a diffuse reflection color vector based on the initial diffuse reflection component map, and calculating an initial specular reflection component map; iteratively optimizing the diffuse reflection component and the specular reflection component until convergence, and the obtained diffuse reflection component map is the intrinsic image.

[0010] Preferably, the network of the encoder-decoder structure is a U-Net structure or a variant thereof; the encoder gradually reduces the resolution of the feature map and extracts high-level semantic features through a series of convolutional layers and downsampling layers; the decoder gradually restores the resolution of the feature map through a series of upsampling layers and convolutional layers, and fuses it with the feature map from the skip connection path.

[0011] Preferably, in the skip connection path of the network, the channel attention vector generated by the deep feature map of the encoder is used to weight the shallow feature map of the encoder, specifically including: performing global average pooling on the deep feature map of the encoder to obtain a channel descriptor; inputting the channel descriptor into a multi-layer perceptron consisting of two fully connected layers, and outputting it through a Sigmoid activation function to generate the channel attention vector; multiplying the channel attention vector by the shallow feature map of the encoder channel by channel to obtain the semantically guided texture feature map.

[0012] Preferably, the unary potential function φ(li) of the conditional random field model constructed by the conditional random field is determined by the preliminary coking area probability map P(li), and its calculation formula is: , where li is the label of pixel i.

[0013] Preferably, the binary potential function ψ(li, lj) is expressed as a weighted sum of two Gaussian kernel functions: , where li and lj are the labels of adjacent pixels i and j, pi and pj are the pixel positions, Ii and Ij are the pixel intensities or color vectors of the intrinsic image, w1, w2, 、 、 are model parameters.

[0014] Preferably, each independent connected area in the refined coking area segmentation map is post-processed, specifically including: calculating the area of ​​each independent connected area and the average probability value of the pixels in the area in the preliminary coking area probability map; the connected area with an area smaller than a preset minimum area threshold, or the connected area with an average probability value lower than a preset minimum probability threshold, is determined as a pseudo-coking area and is eliminated.

[0015] In addition, the present invention also proposes an image recognition system for the coking area on the inner wall of a milk powder drying tower, comprising the following modules:

[0016] An intrinsic image determination unit is configured to obtain an image of the inner wall of the drying tower to be identified; based on a preset three-dimensional geometric model of the inner wall of the drying tower, perform a cylindrical coordinate inverse projection transformation on the inner wall image to obtain a planar unfolded image; utilize a physically based bidirectional reflectance distribution function (BRDF) model to decompose the planar unfolded image into a specular reflection component and a diffuse reflection component, and eliminate the specular reflection component to obtain a diffuse reflection component image serving as the intrinsic image;

[0017] a preliminary segmentation unit, configured to input the intrinsic image into an encoder-decoder network for preliminary segmentation, wherein, in a skip connection path of the network, a channel attention vector generated by a deep feature map of the encoder is used to weight a shallow feature map of the encoder to generate a semantically guided texture feature map, and the semantically guided texture feature map is fused with a feature map of a corresponding layer of the decoder, and the network outputs a preliminary focused area probability map;

[0018] a fine segmentation unit, configured to construct a conditional random field model to perform fine segmentation on the probability map, wherein a unary potential function of the conditional random field model is determined by the probability that each pixel in the probability map belongs to the focus area, and a binary potential function combines the label smoothness constraint between adjacent pixels and the inter-pixel distance constraint calculated based on the intrinsic image, thereby obtaining a refined focus area segmentation map by minimizing the energy function of the conditional random field model;

[0019] The identification unit is used to post-process each independent connected area in the refined coking area segmentation map, screen the independent connected areas according to the preset geometric feature threshold and regional average probability value of each connected area, eliminate pseudo coking areas that do not meet the physical characteristics of coking, and obtain the identification result of the coking area.

[0020] The present invention eliminates the geometric distortion of the image caused by the tower wall curve through the cylindrical coordinate reverse projection transformation, decomposes the image using the physics-based BRDF model, eliminates the strong mirror reflection interference of the stainless steel inner wall, and enables the network to identify based on the intrinsic image that reflects the inherent properties of the object, solving the problem that the highlight area is easily misjudged. In the segmentation network, by introducing the channel attention mechanism guided by deep semantics, the attention to the shallow texture features related to the focused area is strengthened, and the model's ability to identify early and weak focused targets is improved. In addition, the construction of a conditional random field model combined with distance constraints can fine-tune the edges of the network's preliminary segmentation results, so that the segmentation contour is closer to the real boundary of the focus, and through multi-dimensional screening of the connected areas, pseudo targets that do not conform to physical properties are filtered out, further ensuring the accuracy and reliability of the final recognition results. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is a flow chart of a specific embodiment;

[0022] Figure 2 This is a schematic diagram of the tower inner wall image acquisition;

[0023] Figure 3 Schematic diagram of cylindrical coordinate reverse projection exchange;

[0024] Figure 4 This is a schematic diagram of image decomposition based on physical models;

[0025] Figure 5 Schematic diagram of refined segmentation of the conditional random field model. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solutions and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with the specific embodiments of this application and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0027] In one embodiment, the present application proposes a method for image recognition of the coking area on the inner wall of a milk powder drying tower, such as Figure 1 As shown, the following steps are included:

[0028] S1, obtaining an image of the inner wall of a drying tower to be identified; performing a cylindrical coordinate inverse projection transformation on the inner wall image based on a preset three-dimensional geometric model of the inner wall of the drying tower to obtain a planar unfolded image; utilizing a physically based bidirectional reflectance distribution function (BRDF) model, decomposing the planar unfolded image into a specular reflection component and a diffuse reflection component, and eliminating the specular reflection component to obtain a diffuse reflection component map as an intrinsic image;

[0029] An industrial camera with a fisheye lens is installed at the center axis of the top of the drying tower, supplemented by a ring LED array for fill light, to collect a panoramic image of the tower wall with a resolution of 2048×2048 pixels, such as Figure 2 shown.

[0030] According to the design drawings of the drying tower, a cylindrical three-dimensional model of its inner wall is established, and the cylindrical three-dimensional model has a certain radius R and height H. The camera is calibrated to obtain its internal parameters such as focal length and principal point and external parameters such as position and posture in the coordinate system inside the tower. For each pixel point in the target plane expansion image, its corresponding three-dimensional spatial coordinates on the cylindrical three-dimensional model are calculated according to its row and column numbers in the expansion image. The camera projection model and the calibrated internal and external parameters are used to reversely calculate the projected pixel position of the three-dimensional space point on the original fisheye image, and the grayscale value of the original image pixel point is assigned to the pixel point of the current plane expansion image. All pixel points are traversed to complete the transformation, such as Figure 3 shown.

[0031] This method, based on the dichromatic reflection model proposed by Tominaga et al., uses the bidirectional reflectance distribution function (BRDF) model, which assumes that the RGB values ​​of pixels in an image are linear combinations of the object's surface color and the light source's color. The planar unwrapped image is converted from RGB color space to luv color space, where the diffuse reflection component is primarily concentrated in a single cluster, while the specular reflection component forms a strip-shaped cluster that leans toward the light source's color. By analyzing the distribution of pixels on a chromaticity diagram, the pixel clusters representing specular reflection are identified and separated, and then removed or suppressed from the original image. The remaining portion is the diffuse reflection component image, which is primarily determined by the object's inherent properties.

[0032] S2, inputting the intrinsic image into an encoder-decoder network for preliminary segmentation, wherein, in the skip connection path of the network, the channel attention vector generated by the deep feature map of the encoder is used to weight the shallow feature map of the encoder to generate a semantically guided texture feature map, and the semantically guided texture feature map is fused with the feature map of the corresponding layer of the decoder, and the network outputs a preliminary focused area probability map;

[0033] The encoder-decoder network structure uses U-Net as the basis. The encoder part uses a pre-trained ResNet34 network to extract multi-level feature maps. In the skip connection, taking the feature map of the deepest layer of the encoder as an example, it is first compressed into a feature vector through a global average pooling operation. After the feature vector passes through two fully connected layers and a Sigmoid activation function, a channel attention weight vector is generated. The channel attention weight vector is multiplied at the channel level with the feature map of the shallow layer of the encoder, such as the second layer, and each channel of the shallow feature map is reweighted to obtain a semantically guided texture feature map. This weighted feature map is spliced ​​and fused with the upsampled feature map of the corresponding second layer of the decoder. The last layer of the network uses a Sigmoid activation function to output a single-channel image of the same size as the input. The value of each pixel in the image represents the probability that the point belongs to the focused area.

[0034] S3, constructing a conditional random field model to perform fine segmentation on the probability map, wherein the unary potential function of the conditional random field model is determined by the probability that each pixel in the probability map belongs to the focus area, and the binary potential function combines the label smoothness constraint between adjacent pixels and the inter-pixel distance constraint calculated based on the intrinsic image, and obtains a refined focus area segmentation map by minimizing the energy function of the conditional random field model;

[0035] A fully connected conditional random field model is used for optimization. The unary potential function of the conditional random field model is calculated based on the probability map output by the network. For any pixel, the potential energy of being divided into coke is the negative logarithm of its probability value. The binary potential function of the model contains two terms. The first term is a smoothing term, which imposes a fixed penalty value when assigning different labels to adjacent pixels; the second term is an appearance term. In one embodiment, the weight of this term is determined by the geodesic distance between pixels. This distance is calculated on the intrinsic image, which combines the spatial Euclidean distance and grayscale value difference of the pixels, so that pixels with similar grayscale are more closely connected. Through an efficient mean field approximate inference algorithm, the minimum value of the energy function is iteratively solved, and finally a segmentation map of the focus area with sharper edges and better details is obtained.

[0036] S4, post-processing each independent connected area in the refined coking area segmentation map, screening the independent connected areas according to the preset geometric feature threshold and regional average probability value of each connected area, eliminating pseudo coking areas that do not meet the physical characteristics of coking, and obtaining the recognition result of the coking area.

[0037] Using the eight-connected region analysis algorithm, all white pixel areas in the refined segmentation map are marked to identify independent connected regions. For each connected region, its area, aspect ratio, and average probability value are calculated in turn. The area threshold is set to 100 pixels, the aspect ratio range is between 0.2 and 5, and the regional average probability threshold is 0.8. Each connected region is checked one by one. Only if its area is greater than 100 pixels, the aspect ratio is between 0.2 and 5, and the average value of all its internal pixels in the probability map output by the network is greater than 0.8, the region is retained. Otherwise, it is regarded as a pseudo-coking region caused by welds or noise and is removed from the segmentation map.

[0038] In a feasible embodiment, the step of decomposing the planar expanded image into a specular reflection component and a diffuse reflection component using a physics-based bidirectional reflectance distribution function (BRDF) model specifically includes: using a two-color reflection model to decompose the pixel value I of any pixel point in the planar expanded image into a linear superposition of a diffuse reflection component M and a specular reflection component S; obtaining an initial diffuse reflection component map by calculating the minimum color channel value of the image and performing a morphological closing operation on the channel image; estimating a diffuse reflection color vector based on the initial diffuse reflection component map, and calculating an initial specular reflection component map; iteratively optimizing the diffuse reflection component and the specular reflection component until convergence, and the obtained diffuse reflection component map is the intrinsic image.

[0039] The decomposition method of the plane unfolded image into the specular reflection component and the diffuse reflection component is based on a physical assumption that the reflected light on the surface of the object consists of diffuse reflection and specular reflection. Figure 4 As shown in Figure 2. Diffuse reflection is related to the intrinsic color and texture of an object and is light that is scattered evenly in all directions. Specular reflection is a highlight formed by direct reflection of a light source in a specific direction and typically takes on the color of the light source, such as white. In an image of a coking chamber, the color and texture of the coke itself are part of the diffuse reflection component, while bright spots caused by surface unevenness or moisture are part of the specular reflection component. Removing the specular reflection component is crucial for accurately identifying the true areas of coking.

[0040] The execution process of the decomposition method is as follows: for the red, green and blue color values ​​of an input pixel, for example, 180, 160, 150, the minimum value 150 is taken as the initial diffuse reflection intensity. After this operation is performed on the entire image, a morphological closing operation is performed using a structure element of, for example, 5×5 size to fill the small holes that may be generated by highlight removal, and a smooth initial diffuse reflection component map is obtained. Based on this map, the average color of the coke area is estimated, for example, 110, 90, 80, as the diffuse reflection color vector. By subtracting the diffuse reflection component from the original image, the initial specular reflection component map is obtained. Through an iterative algorithm, for example, the number of iterations is set to 50 times or the image difference between two iterations is less than one ten-thousandth, the diffuse reflection and specular reflection components are alternately updated and optimized until both reach a stable state. The final diffuse reflection component map is the intrinsic image without highlight interference.

[0041] In a feasible embodiment, the network of the encoder-decoder structure is a U-Net structure or a variant thereof; the encoder gradually reduces the resolution of the feature map and extracts high-level semantic features through a series of convolutional layers and downsampling layers; the decoder gradually restores the resolution of the feature map through a series of upsampling layers and convolutional layers, and fuses it with the feature map from the skip connection path.

[0042] The U-Net architecture was chosen for its exceptional performance in pixel-level segmentation tasks, and is particularly well-suited for objects with irregular boundaries and complex textures, such as coke deposits. The left half of the U-Net is the encoder, responsible for feature extraction, while the right half is the decoder, responsible for accurately reproducing the segmentation results. The encoder mimics an image classification network, using continuous convolution and pooling operations to gradually compress the input image from, for example, 256 by 256 pixels to 16 by 16 pixels, increasing the number of feature channels from 64 to 1024. This process captures the image's deep semantic information, specifically understanding which areas of the image, at a macro level, resemble coke deposits.

[0043] The decoder performs the opposite operation of the encoder. It gradually restores the 16×16 condensed feature map to its original size of 256 by 256 by upsampling, such as using transposed convolution. The key is that after each upsampling, the decoder will concatenate and fuse the obtained feature map with the feature map of the corresponding level of the encoder through jump connections. For example, when the decoder restores the feature map to 32×32 size, it is fused with the feature map of the encoder, which is also 32×32 in size. This allows the network to simultaneously utilize high-level semantic information on the decoder path and low-level texture and edge detail information on the encoder path to generate a segmentation map of the focused area with clear boundaries and precise positioning.

[0044] In a feasible embodiment, in the skip connection path of the network, the channel attention vector generated by the deep feature map of the encoder is used to weight the shallow feature map of the encoder, specifically including: performing global average pooling on the deep feature map of the encoder to obtain a channel descriptor; inputting the channel descriptor into a multi-layer perceptron consisting of two fully connected layers, and outputting it through a Sigmoid activation function to generate the channel attention vector; and multiplying the channel attention vector by the shallow feature map of the encoder channel by channel to obtain the semantically guided texture feature map.

[0045] In the U-Net's skip connections, the shallow feature maps passed from the encoder to the decoder contain rich spatial details and textures, but they also contain a large amount of background texture unrelated to focus. To address this issue, feature maps from deeper layers of the encoder are utilized. These feature maps contain higher-level semantic information and are capable of providing a preliminary understanding of image content. For example, a deep feature map of size 16×16×512 is considered to contain global contextual information about possible focus locations.

[0046] The deep feature map is compressed into a 512-dimensional vector through global average pooling. This vector can be regarded as the summary information of 512 feature channels. The vector passes through a multi-layer perceptron containing two fully connected layers. For example, the first layer compresses the dimension from 512 to 32, and the second layer restores it back to 512 to learn the complex dependencies between channels. Finally, an activation function is used to output a 512-dimensional attention vector, in which the value of each element is between 0 and 1, representing the importance of the corresponding channel. By multiplying this attention vector with the shallow feature map of the encoder, such as a 128×128×512, channel by channel, the weights of the texture feature channels related to the focus semantics can be amplified, and irrelevant noise feature channels can be suppressed, thus delivering semantically filtered, higher-quality texture information to the decoder.

[0047] In a feasible embodiment, the unary potential function φ(li) of the conditional random field model constructed by the conditional random field is determined by the preliminary coking area probability map P(li), and its calculation formula is: , where li is the label of pixel i, such as Figure 5 shown.

[0048] The conditional random field model is used as a post-processing technique to optimize the segmentation results output by the neural network. The model is based on a unary potential function, which assigns a classification preference to each pixel in the image. This preference is derived entirely from the preliminary probability map of focal regions generated by the neural network in the previous step. Each pixel value on this probability map represents the probability that the pixel belongs to a focal region.

[0049] By taking the negative logarithm of the probability value to calculate the unary potential function, an intuitive cost relationship is established. If the neural network determines that a pixel is coke with an extremely high probability, such as 0.98, then the corresponding unary potential function value, that is, the negative logarithm 0.98, will be very close to 0. This means that the cost of marking the pixel as coke is extremely low. Conversely, if the probability is very low, such as 0.01, its negative logarithm value will be very large, which means that the cost of marking the pixel as coke is extremely high. The unary potential function converts the neural network's prediction results into an initial per-pixel labeling cost that the conditional random field model can understand.

[0050] In a feasible embodiment, the binary potential function ψ(li, lj) is expressed as a weighted sum of two Gaussian kernel functions: , where li and lj are the labels of adjacent pixels i and j, pi and pj are the pixel positions, Ii and Ij are the pixel intensities or color vectors of the intrinsic image, w1, w2, 、 、 are model parameters.

[0051] The binary potential function plays a crucial role in the conditional random field model. It defines the cost of the label relationship between adjacent pixels, aiming to make the boundary of the segmentation result smoother and more consistent with the true edge of the image. The binary potential function encourages adjacent pixels with similar appearance to have the same label, that is, to belong to the same focus or background, while allowing adjacent pixels with significantly different appearances to have different labels. Using the intrinsic image obtained in the previous step rather than the original image to calculate the pixel intensity difference can avoid the interference of lighting and shadow changes on edge judgment, allowing the model to focus more on the color and texture boundaries of the object itself.

[0052] The binary potential function consists of two weighted parts. The first part is a smoothing kernel, which only considers the spatial distance between pixels. Regardless of the pixel color, as long as two pixels are close, they tend to have the same label, which helps to eliminate isolated noise points. The second part is an appearance kernel, which considers the spatial distance and color difference of pixels. For example, for two adjacent pixels, if they are close in space and their colors in the intrinsic image are also very close, then the cost of assigning them different labels will be very high. But if their colors are very different, such as one on the coke and the other on the furnace wall, then even if they are close together, the cost of assigning different labels will be small. By adjusting the weights and parameters of the two Gaussian kernels, the balance between maintaining the internal consistency of the region and respecting the boundaries of the real object can be finely controlled.

[0053] In a feasible embodiment, each independent connected area in the refined coking area segmentation map is post-processed, specifically including: calculating the area of ​​each independent connected area and the average probability value of the pixels in the area in the preliminary coking area probability map; the connected areas with an area smaller than a preset minimum area threshold, or the connected areas with an average probability value lower than a preset minimum probability threshold, are determined as pseudo coking areas and are eliminated.

[0054] To remove areas most likely to be false positives, the segmentation map may still contain small, isolated spots that are mistakenly identified as coke after processing with the neural network and conditional random fields, or larger areas where the network's overall confidence is low. These are usually artifacts caused by image noise, reflections, or similar textures on the furnace wall.

[0055] The post-processing method sets two screening criteria. All independent areas marked as coking on the segmentation map are analyzed. The first criterion is area. For example, a minimum area threshold is preset to 100 pixels. Any coking area smaller than this area will be directly regarded as noise and removed from the result. The second criterion is the average confidence. For areas that pass the area screening, the probability average of all pixels constituting the area on the original output probability map of the neural network is calculated. If this average probability is lower than a preset threshold, such as 0.80, it means that the network is not confident enough to judge this area as coking as a whole. Such an area may be an artifact even if it is large in area, so it will also be eliminated. Through this dual-standard filtering, the accuracy and reliability of the final coking detection results can be improved.

[0056] It should be noted that for the aforementioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the embodiments of this specification are not limited by the order of the actions described, because according to the embodiments of this specification, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the embodiments of this specification.

[0057] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0058] The preferred embodiments disclosed above are intended only to help illustrate this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific embodiments described. Obviously, many modifications and variations can be made based on the content of the embodiments of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the embodiments of this specification, so that those skilled in the art can better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A method for identifying an image of a coking area on the inner wall of a milk powder drying tower, characterized in that: The following steps are involved: Acquire an image of the inner wall of the drying tower to be identified; perform a cylindrical coordinate inverse projection transformation on the inner wall image based on a preset three-dimensional geometric model of the inner wall of the drying tower to obtain a planar unfolded image; utilize a physics-based bidirectional reflectance distribution function (BRDF) model to decompose the planar unfolded image into a specular reflection component and a diffuse reflection component, and eliminate the specular reflection component to obtain a diffuse reflection component map as an intrinsic image; The intrinsic image is input into an encoder-decoder network for preliminary segmentation, wherein, in the skip connection path of the network, the channel attention vector generated by the deep feature map of the encoder is used to weight the shallow feature map of the encoder to generate a semantically guided texture feature map, and the semantically guided texture feature map is fused with the feature map of the corresponding layer of the decoder, and the network outputs a preliminary focus area probability map; A conditional random field model is constructed to perform fine segmentation on the probability map, wherein a unary potential function of the conditional random field model is determined by the probability that each pixel in the probability map belongs to the focus area, and a binary potential function combines the label smoothness constraint between adjacent pixels and the inter-pixel distance constraint calculated based on the intrinsic image. A refined focus area segmentation map is obtained by minimizing the energy function of the conditional random field model; The independent connected regions in the refined coking region segmentation map are post-processed, and the independent connected regions are screened according to the preset geometric feature thresholds and regional average probability values ​​of each connected region, and pseudo coking regions that do not conform to the physical characteristics of coking are eliminated to obtain the recognition results of the coking regions.

2. The method according to claim 1, characterized in that The step of decomposing the planar expanded image into a specular reflection component and a diffuse reflection component by using a physics-based bidirectional reflectance distribution function (BRDF) model specifically includes: Using the two-color reflection model, the pixel value I of any pixel point in the plane expanded image is decomposed into the linear superposition of the diffuse reflection component M and the specular reflection component S; The initial diffuse component map is obtained by calculating the minimum color channel value of the image and performing morphological closing operation on the channel image; estimating a diffuse reflection color vector according to the initial diffuse reflection component map, and calculating an initial specular reflection component map; The diffuse reflection component and the specular reflection component are iteratively optimized until convergence, and the obtained diffuse reflection component image is the intrinsic image.

3. The method according to claim 1, characterized in that The network of the encoder-decoder structure is a U-Net structure or a variant thereof; the encoder gradually reduces the resolution of the feature map and extracts high-level semantic features through a series of convolutional layers and downsampling layers; The decoder gradually restores the resolution of the feature map through a series of upsampling layers and convolutional layers, and fuses it with the feature map from the skip connection path.

4. The method according to claim 1, wherein In the skip connection path of the network, the channel attention vector generated by the deep feature map of the encoder is used to weight the shallow feature map of the encoder to generate a semantically guided texture feature map, specifically including: Performing global average pooling on the deep feature map of the encoder to obtain a channel descriptor; The channel descriptor is input into a multi-layer perceptron consisting of two fully connected layers and output through a Sigmoid activation function to generate a channel attention vector; The channel attention vector is multiplied by the shallow feature map of the encoder channel by channel to obtain the semantically guided texture feature map.

5. The method according to claim 1, wherein The unary potential function φ(li) of the conditional random field model is determined by the preliminary coking area probability map P(li), and the calculation formula is: , where li is the label of pixel i.

6. The method according to claim 1, characterized in that The binary potential function ψ(li, lj) is expressed as the weighted sum of two Gaussian kernel functions: , where li and lj are the labels of adjacent pixels i and j, pi and pj are the pixel positions, Ii and Ij are the pixel intensities or color vectors of the intrinsic image, w1, w2, 、 、 are model parameters.

7. The method according to claim 1, characterized in that Post-processing is performed on each independent connected region in the refined coking region segmentation map, specifically including: Calculate the area of ​​each independent connected region and the average probability value of the pixels in the region in the preliminary coking area probability map; Connected areas with an area smaller than a preset minimum area threshold, or connected areas with an average probability value lower than a preset minimum probability threshold, are determined to be pseudo-coking areas and are removed.

8. A system for recognizing the coking area on the inner wall of a milk powder drying tower, characterized in that: Includes the following modules: An intrinsic image determination unit is configured to obtain an image of the inner wall of the drying tower to be identified; based on a preset three-dimensional geometric model of the inner wall of the drying tower, perform a cylindrical coordinate inverse projection transformation on the inner wall image to obtain a planar unfolded image; utilize a physically based bidirectional reflectance distribution function (BRDF) model to decompose the planar unfolded image into a specular reflection component and a diffuse reflection component, and eliminate the specular reflection component to obtain a diffuse reflection component image serving as the intrinsic image; a preliminary segmentation unit, configured to input the intrinsic image into an encoder-decoder network for preliminary segmentation, wherein, in a skip connection path of the network, a channel attention vector generated by a deep feature map of the encoder is used to weight a shallow feature map of the encoder to generate a semantically guided texture feature map, and the semantically guided texture feature map is fused with a feature map of a corresponding layer of the decoder, and the network outputs a preliminary focused area probability map; a fine segmentation unit, configured to construct a conditional random field model to perform fine segmentation on the probability map, wherein a unary potential function of the conditional random field model is determined by the probability that each pixel in the probability map belongs to the focus area, and a binary potential function combines the label smoothness constraint between adjacent pixels and the inter-pixel distance constraint calculated based on the intrinsic image, thereby obtaining a refined focus area segmentation map by minimizing the energy function of the conditional random field model; The identification unit is used to post-process each independent connected area in the refined coking area segmentation map, screen the independent connected areas according to the preset geometric feature threshold and regional average probability value of each connected area, eliminate pseudo coking areas that do not meet the physical characteristics of coking, and obtain the identification result of the coking area.

9. The system according to claim 8, characterized in that The step of decomposing the planar expanded image into a specular reflection component and a diffuse reflection component by using a physics-based bidirectional reflectance distribution function (BRDF) model specifically includes: Using the two-color reflection model, the pixel value I of any pixel point in the plane expanded image is decomposed into the linear superposition of the diffuse reflection component M and the specular reflection component S; The initial diffuse component map is obtained by calculating the minimum color channel value of the image and performing morphological closing operation on the channel image; estimating a diffuse reflection color vector according to the initial diffuse reflection component map, and calculating an initial specular reflection component map; The diffuse reflection component and the specular reflection component are iteratively optimized until convergence, and the obtained diffuse reflection component image is the intrinsic image.

10. The system according to claim 8, wherein: The network of the encoder-decoder structure is a U-Net structure or a variant thereof; the encoder gradually reduces the resolution of the feature map and extracts high-level semantic features through a series of convolutional layers and downsampling layers; The decoder gradually restores the resolution of the feature map through a series of upsampling layers and convolutional layers, and fuses it with the feature map from the skip connection path.

Images