A ladle infrared monitoring image recognition method and device and a storage medium

Abstract

The application discloses a ladle infrared monitoring image recognition method and device and a storage medium, which comprises the following steps: acquiring one or more infrared thermal imaging images of a ladle; performing feature extraction on the acquired infrared thermal imaging images of the ladle by using a full convolutional neural network (FCN) and outputting a first feature image corresponding to a low-resolution infrared thermal imaging image of the ladle and / or a second feature image corresponding to a high-resolution infrared thermal imaging image of the ladle; performing coarse segmentation on the first feature image and / or the second feature image; performing fine segmentation on the first feature image and / or the second feature image after the coarse segmentation by using a Gibbs distribution; and performing performance evaluation on the first feature image and / or the second feature image after the fine segmentation.

Description

Technical Field

[0001] This application relates to the fields of metallurgy and artificial intelligence, and in particular to a method, apparatus and storage medium for recognizing infrared monitoring images of steel ladles based on a fully convolutional neural network. Background Technology

[0002] The metallurgical industry is currently facing challenges such as overcapacity, fierce market competition, increasing environmental pressure, and slowing demand growth. It is crucial to achieve sustainable development by improving technology to ensure safe production, reduce costs and increase efficiency, and strengthen environmental governance.

[0003] As a critical equipment container in the metallurgical process, the steel ladle is used in steel mills and foundries to receive molten steel and perform pouring operations before open-hearth furnaces, electric furnaces, or converters. Due to safety concerns, personnel cannot closely monitor the ladle during molten steel reception and transfer. Furthermore, the residual molten steel in the ladle can cool and affect its identification after use. Therefore, how to achieve reliable monitoring and identification control of the ladle during its use through technological means has been a pressing and key focus for steel mills. Summary of the Invention

[0004] This application was made in view of the above-mentioned problems. The purpose of this application is to provide a steel ladle infrared monitoring image recognition method based on a fully convolutional neural network, so as to improve the recognition accuracy of steel ladle features during the molten iron transfer process.

[0005] According to a first aspect of this application, a method for recognizing infrared monitoring images of steel ladles based on a fully convolutional neural network is provided, comprising:

[0006] Acquire one or more infrared thermal images of the ladle;

[0007] The infrared thermal imaging image of the steel ladle is used to extract features using a fully convolutional neural network (FCN), and outputs a first feature image corresponding to the low-resolution infrared thermal imaging image of the steel ladle, and / or a second feature image corresponding to the high-resolution infrared thermal imaging image of the steel ladle.

[0008] Coarse segmentation is performed on the first feature image and / or the second feature image, wherein in the coarse segmentation process, one-dimensional feature values ​​in the first feature image and / or the second feature image are segmented and features are extracted.

[0009] The first and / or second feature images after the coarse segmentation are further segmented using a Gibbs distribution. In this fine segmentation, the first and / or second feature images after the coarse segmentation are correlated and their features extracted by considering the context of the infrared thermal imaging image of the ladle.

[0010] The performance of the first feature image and / or the second feature image after fine segmentation is evaluated.

[0011] In a second aspect of this application, a ladle infrared monitoring image recognition device is provided, comprising: a memory and a processor; the memory stores a computer program, which, when executed by the processor, represents the aforementioned ladle infrared monitoring image recognition method.

[0012] In a third aspect of this application, a computer-readable storage medium is provided, having stored thereon computer program instructions, characterized in that the computer program instructions, when executed by a processor, implement the aforementioned ladle infrared monitoring image recognition method.

[0013] Invention Effects

[0014] In this application, a fully convolutional neural network-dense conditional random field (FCN-DCRF) deep learning segmentation algorithm is used to perform coarse and fine segmentation identification on infrared thermal imaging images of steel ladles during use. This effectively improves the automatic identification performance of features such as the boom, ladle opening, ladle bottom, and ladle number, and enhances the accuracy of image feature recognition. By monitoring the ladle status in real time, early warnings are issued for abnormal conditions, improving the safety of the ladle during molten steel receiving and transfer. Attached image description:

[0015] Figure 1 FCN-DCRF Image Convolutional Deep Learning Process According to Embodiments of this Application

[0016] Figure 2 This is an example of an infrared thermal imaging image of a steel ladle obtained according to an embodiment of this application.

[0017] Figure 3 These are parameters used to evaluate the infrared thermal imaging image recognition performance of a steel ladle according to embodiments of this application.

[0018] Figure 4 This is an example of identifying effective infrared thermal imaging images of different locations of a ladle according to an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0020] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0021] like Figure 1 As shown, this application discloses a method for recognizing infrared monitoring images of steel ladles based on a fully convolutional neural network, including the following steps:

[0022] S11: Perform infrared detection on the ladle to acquire one or more infrared thermal imaging images of the ladle.

[0023] In one implementation, the infrared thermal imaging device is supported by a suspension support, suspended from the ceiling. Furthermore, the infrared thermal imaging device is supported by the suspension support in a manner capable of movement along at least one of the X, Y, and Z axes. The suspension support supports the infrared thermal imaging device in a movable and rotatable manner. The infrared thermal imaging device includes one or more infrared imagers / cameras that acquire infrared thermal images of the ladle in real time. An operator can remotely control the movement and shooting direction of the infrared thermal imaging device via a remote control.

[0024] In one implementation, one or more infrared imagers / cameras mounted on a ladle car are used to acquire infrared thermal images of the ladle in real time.

[0025] The infrared thermal imaging images obtained in this application may be static images, single images, or independent images, or may be images included in a series of moving images, such as frames in a video sequence of video frames, or images in a sequence or collection of images of different types.

[0026] In one implementation, the infrared thermal imaging device is positioned at a predetermined distance from the ladle to obtain a lower temperature environment relative to the molten steel flow itself, reducing the possibility of damage to the infrared thermal imaging unit due to proximity to the ladle, and is easy to operate.

[0027] In one implementation, an infrared imager / camera is used to capture infrared thermal images of the ladle at predetermined intervals and frame rates, while monitoring for anomalies in the infrared thermal images of the ladle and performing the operation.

[0028] Steel ladles can be detected using different infrared wavelengths of an infrared imager / camera. In a preferred embodiment, a thermal imaging image of the steel ladle is detected using a far-infrared wavelength (greater than or equal to 8 μm).

[0029] Figure 2 This illustration shows an example of how this application utilizes an infrared imager / camera to acquire real-time infrared thermal images of a steel ladle.

[0030] The infrared thermal images of the ladle captured in this way can be stored on one or more storage devices implementing the system, or on storage devices accessible to the system, such as connected storage devices, like local storage devices and / or storage devices connected via a network. In various embodiments, the infrared thermal images of the ladle can be obtained from various sources. For a particular user, the infrared thermal images of the ladle can be obtained from uploads made by the user to a server via one or more networks, from an album of multiple images owned or accessible by the user, or from other storage collections, etc.

[0031] S12: Feature extraction is performed on the infrared thermal imaging image of the steel ladle using a fully convolutional neural network (FCN) model;

[0032] In one possible implementation, optionally, preprocessing is performed on one or more infrared thermal imaging images of the acquired ladle. The preprocessing includes screening the acquired infrared thermal imaging images of the ladle, removing unqualified data such as image blurring caused by imaging device instability or unknown reasons, and generating an image dataset that meets the requirements for feature extraction.

[0033] The preprocessing also includes: for infrared thermal imaging images with blurred edges and features, preprocessing such as edge pixel supplementation, image denoising, and / or enhancement is performed. Image denoising includes filtering the infrared thermal imaging image of the ladle using mean filtering, median filtering, or a denoising autoencoder, etc. Image enhancement includes using histogram equalization, adaptive histogram equalization, Fourier transform, etc., to enhance the contrast of the infrared thermal imaging image of the ladle.

[0034] Preprocessing the acquired infrared thermal image of the ladle also includes annotating the image with semantic tags to distinguish between target and background areas. This can be achieved by assigning different binary values ​​to the pixels of the ladle's infrared thermal image using an autoencoder.

[0035] Feature extraction is performed on the raw or pre-processed infrared thermal imaging images of the ladle using a fully convolutional neural network (FCN) model. The specific steps include:

[0036] (1) Receive the original or preprocessed infrared thermal imaging image of the ladle that meets the network input requirements through the input layer of the fully convolutional neural network (FCN) model;

[0037] (2) Feature extraction is performed on the infrared thermal imaging image of the received ladle using the convolutional layers of the fully convolutional neural network (FCN) model. Convolution refers to the process of multiplying the input image matrix and the convolution kernel matrix by weights, and then passing the result through an activation function to output a new feature map.

[0038] When performing image feature extraction on infrared thermal imaging images of steel ladles, the first step is to detect the sharpness of the input, such as the pre-processed infrared thermal imaging image of the steel ladle. Many factors can cause image blurring during the acquisition, transmission, and processing of infrared thermal imaging images. For example, incorrect focusing of the infrared thermal imaging device on the steel ladle during image acquisition can cause defocus blur; relative motion of the steel ladle relative to the infrared thermal imaging device can cause motion blur; and high-frequency loss after compression of the infrared thermal imaging image of the steel ladle can cause blurring. Blur reduces the sharpness of the infrared thermal imaging image of the steel ladle, seriously affecting the image quality and leading to difficulties or even failures in feature extraction. Therefore, it is necessary to use effective blur evaluation methods to control the use of blurred images in the infrared thermal imaging image of the steel ladle, thereby improving the recognition performance of the infrared thermal imaging image of the steel ladle.

[0039] In this application, the algorithm for detecting the sharpness of the infrared thermal imaging image of a ladle includes: (1) a pixel-based technique, which includes analyzing the statistical characteristics of the pixel grayscale values ​​of the infrared thermal imaging image of the ladle and the correlation between pixels; (2) a transform domain-based technique, which utilizes the principle that the more high-frequency components in the transform domain, the clearer the infrared thermal imaging image of the ladle, and the fewer high-frequency components, the blurrier the infrared thermal imaging image of the ladle; and (3) an image gradient-based technique, which uses the gradient of the image edge to measure the image blurriness of the infrared thermal imaging image of the ladle. The larger the gradient, the clearer the infrared thermal imaging image of the ladle.

[0040] In one possible implementation, if the resolution of the detected infrared thermal imaging image of the ladle is determined to be lower than a predetermined threshold, then the infrared thermal imaging image of the ladle below the predetermined threshold is identified as a low-resolution infrared thermal imaging image of the ladle, i.e., a first-resolution infrared thermal imaging image of the ladle, and the low-resolution infrared thermal imaging image is subjected to upsampling / deconvolution processing using an upsampling block; if the resolution of the detected infrared thermal imaging image of the ladle is determined to be equal to or greater than the predetermined threshold, then the infrared thermal imaging image of the ladle equal to or greater than the predetermined threshold is identified as a high-resolution infrared thermal imaging image of the ladle, i.e., a second-resolution infrared thermal imaging image of the ladle, and the high-resolution infrared thermal imaging image is subjected to downsampling / convolution processing using a downsampling block.

[0041] In one possible implementation, the fully convolutional neural network (FCN) model of this application performs upsampling / deconvolution processing on the first resolution steel ladle infrared thermal imaging image using the following formula (1):

[0042] N=(w-1)×s+k-2p (1)

[0043] Where N is the size of the output first-resolution steel ladle infrared thermal image, w is the size of the input first-resolution steel ladle infrared thermal image, s is the convolution stride, k is the convolution kernel size, and p is the supplementary edge pixels.

[0044] In this application, the fully convolutional neural network (FCN) model performs upsampling / deconvolution on the first-resolution steel ladle infrared thermal imaging image, and performs image interpolation and other processing to ensure that the network output image of the first-resolution steel ladle infrared thermal imaging image is consistent with the size of the input first-resolution steel ladle infrared thermal imaging image, so as to ensure that image features can be reliably extracted, and by enlarging the original infrared thermal imaging image, it can be displayed on a higher resolution device.

[0045] The fully convolutional neural network (FCN) model performs downsampling / convolution processing on the second-resolution infrared thermal imaging image of the ladle using the following formula (2):

[0046] N=(W-k+2Q) / S+1 (2)

[0047] The output second-resolution steel ladle infrared thermal imaging image has a size of N×N, the input second-resolution steel ladle infrared thermal imaging image has a size of W×W, the filter size is k×k, the step size is S, and the supplementary edge pixels are Q.

[0048] The fully convolutional neural network (FCN) model downsamples the second-resolution steel ladle infrared thermal imaging image by downsampling / convolution, reducing the image size so that the second-resolution steel ladle infrared thermal imaging image conforms to the display area size, and generating a thumbnail of the corresponding infrared thermal imaging image, thereby reducing the dimensionality and workload of convolution and image extraction.

[0049] In image processing using a fully convolutional neural network (FCN) model, the activation function performs a nonlinear transformation on the convolved first-resolution and / or second-resolution infrared thermal imaging images of a ladle, thereby enhancing the expressive power of the FCN. This application uses a rectified linear unit (ReLU) as the activation function. The expression for the ReLU function is as follows: ReLU(y) = max(0,y), where y is the input value and ReLU(y) is the output value. If the input value y is greater than 0, the output value is y; if the input value y is less than or equal to 0, the output value is 0.

[0050] (3) The feature dimension of the first-resolution and / or second-resolution ladle infrared thermal imaging images is reduced by using the pooling layer of the fully convolutional neural network (FCN) model, thereby reducing the number of model parameters and computational cost. The pooling operations performed by the pooling layer in this application include average pooling or max pooling.

[0051] Assuming the input first-resolution ladle infrared thermal imaging image and / or second-resolution ladle infrared thermal imaging image are 4×4 in size, the sampling step size of the pooling layer is 2, the padding is 0, and the pooling kernel size is 2×2, then after pooling, the output first-resolution ladle infrared thermal imaging image and / or second-resolution ladle infrared thermal imaging image will be 2×2 in size.

[0052] (4) The output layer of the fully convolutional neural network (FCN) model is used to output the first feature image after performing upsampling / deconvolution processing on the first resolution steel ladle infrared thermal imaging image by formula (1), or the second feature image after performing downsampling / convolution processing on the second resolution steel ladle infrared thermal imaging image by formula (2).

[0053] In this application, the results of feature extraction from infrared thermal imaging images of steel ladles at different resolutions include feature images corresponding to the bottom, opening, and lifting rings of the steel ladle, as well as temperature image features corresponding to different areas of the steel ladle carrying molten steel.

[0054] S13; Perform coarse segmentation processing on the first feature image and / or the second feature image output in step S12.

[0055] Coarse segmentation refers to the operation of segmenting and extracting features from the first feature image and / or second feature image obtained after convolution, pooling, and full convolution of the infrared thermal imaging image of the ladle by a fully convolutional neural network.

[0056] In one possible approach, a region of interest (ROI) is defined or segmented from the coarsely segmented infrared thermal image of the ladle by combining one-dimensional feature values ​​and semantics.

[0057] In the semantic segmentation result image obtained by performing coarse segmentation, the pixel value of the region of interest in the infrared thermal image of the ladle is 1, and the pixel value of the remaining regions is 0.

[0058] The coarse segmentation process enables pixel-level classification of the infrared thermal imaging image of the ladle, which can effectively achieve ROI segmentation.

[0059] S14; Perform fine segmentation on the first feature image and / or the second feature image after coarse segmentation.

[0060] In the above coarse segmentation process, only key features are extracted, without fully combining the context of the steel ladle's infrared thermal imaging image (i.e., the surrounding area adjacent to the coarsely segmented steel ladle's infrared thermal imaging image) for correlation identification and feature extraction. This results in rough segmentation edges, leading to inaccurate image segmentation and low precision.

[0061] To improve feature extraction, this application performs fine segmentation on the first and / or second feature images after coarse segmentation, which can further improve the segmentation accuracy of the infrared thermal imaging image of the ladle.

[0062] This application performs fine segmentation on the first and / or second feature images of the ladle after coarse segmentation using the Gibbs distribution, specifically including:

[0063] For a coarsely segmented first feature image and / or second feature image of size W×H, let I be the observation vector of the entire first feature image and / or second feature image, and X be the class label corresponding to all observation vectors. Then the conditional random field of the coarsely segmented first feature image and / or second feature image can be represented by the Gibbs distribution as follows:

[0064]

[0065] Where W is the width of the first feature image and / or the second feature image after coarse segmentation; H is the height of the first feature image and / or the second feature image after coarse segmentation; in formula (3), Z(I) is the normalization factor, which can be understood as a constant variable.

[0066] E(X|I) is the energy function:

[0067] Among them Ψ i (x i )=-logP(x i |I), is the unary potential energy of the probability output of the last layer of the fully convolutional neural network (FCN) model; Ψ ij (x i ,x jΨ refers to the binary potential function of a fully convolutional neural network (FCN) model, which describes the spatial context relationship between pixels i and j in the first and / or second feature images after coarse segmentation. ij (x i ,X j ) is defined as:

[0068]

[0069] In equation (5) above, the values ​​of i and j range from 1 to N (N = W × H); p i and I j Representing the coordinates and intensity of pixel i in the original infrared thermal image of the ladle, respectively, |p i -p j |Refers to the distance between pixel i and pixel j in the original infrared thermal image of the ladle,|I i -I j |Refers to the difference in grayscale values ​​between pixels in the original infrared thermal image of the ladle; ω (1) and ω (2) The weights of a fully convolutional neural network (FCN) model; θ α θ β θ β It is the standard deviation parameter of the Gaussian kernel.

[0070] By using the Gibbs distribution to refine the coarsely segmented first and / or second feature images, and then associating the features of the obtained first and / or second feature images with the context of the original infrared thermal image of the ladle, more accurate image segmentation features can be obtained.

[0071] S14: Evaluate the performance of the first and / or second feature images after fine segmentation.

[0072] This application employs four evaluation metrics to assess the performance of the FCN algorithm in performing the aforementioned processing on the infrared thermal imaging image of a steel ladle: Precision (P), Recall (R), and F... β The index (F-measure, F) β The four indices are defined as follows: (The text then lists the indices and their definitions.)

[0073]

[0074]

[0075]

[0076]

[0077] In the above formulas (6)-(9), S1 represents the set of correct pixels in the first and / or second feature images after fine segmentation extracted by the FCN algorithm; S2 represents the set of all pixels in the first and / or second feature images after fine segmentation extracted by the FCN algorithm; S3 represents the set of all pixels in the labeled area of ​​the infrared thermal imaging image of the steel ladle set manually, and the operator |.| represents the statistics of pixels in this area; F β The F1 index is a weighted harmonic average of recall and precision. When the confidence level β = 1, it becomes the most common F1 index. The closer the F1 index is to 1, the better the segmentation performance. The J index is used to measure the region similarity between the FCN algorithm segmentation and the manual segmentation results. Similarly, the closer the J index is to 1, the better the segmentation performance.

[0078] Figure 3 This illustrates the evaluation level set using the above four indicators when evaluating the finely segmented first feature image and / or second feature image in one embodiment of this application. The accuracy (P), recall (R), and accuracy (F) of the infrared thermal imaging image of the ladle are also considered. β When the index and J index meet the "low" level, it is determined that there is a feature recognition difference between the first and / or second feature images after fine segmentation, and the FCN algorithm needs to be used again to iterate the first and / or second feature images after fine segmentation. When the precision (P), recall (R), and F index of the first and / or second feature images after fine segmentation are... β When the index and J index meet the level of "medium" or "high", it is determined that the feature recognition of the first feature image and / or the second feature image after fine segmentation is good or high, which meets the requirements for the recognition and monitoring of infrared thermal imaging images of steel ladles.

[0079] In one specific embodiment, this application annotates the original infrared thermal imaging image of the steel ladle with labels containing binary semantic information, namely ROI and background; performs image feature extraction on the original infrared thermal imaging image of the steel ladle using an FCN model; inputs the infrared thermal imaging image of the steel ladle after feature extraction through a fully convolutional neural network (FCN) into a coarse segmentation module for coarse segmentation processing; inputs the first feature image and / or the second feature image after coarse segmentation into a fine segmentation module; and calculates pixel potential energy using the spatial context information of the image region to further obtain the recognition results of the final fine segmented first feature image and / or second feature image.

[0080] Figure 4An example is shown that the evaluation index given by the above formulas (6)-(9) is used to evaluate the first feature image and / or the second feature image after fine segmentation according to the present application, and outputs the first feature image and / or the second feature image after fine segmentation that meets the specified image processing conditions.

[0081] In this application, for the first feature image and / or the second feature image after fine segmentation that does not meet the image processing requirements, they are returned to steps S13 and S14 for coarse segmentation and fine segmentation iterative processing.

[0082] In one implementation, preferably, this application further performs deep learning on the finely segmented first feature image and / or second feature image that meet the performance evaluation indicators.

[0083] In deep learning, based on the feature values ​​extracted from the finely segmented first and / or second feature images that meet the performance evaluation criteria, a small-batch, high-frequency deep learning training process is performed to quickly identify different regions in the infrared thermal imaging image of a steel ladle.

[0084] By utilizing the steel ladle infrared monitoring image recognition method based on fully convolutional neural networks in this application, it is possible to identify image features of the steel ladle boom, ladle opening, ladle bottom, and ladle body under different postures. By combining the relationship between the image features and the context associated with the identified image features, image features can be identified more accurately, thereby improving recognition efficiency.

[0085] In one example, this application identifies temperature features by recognizing the temperature characteristics of images of the ladle opening, body, and bottom in different postures corresponding to molten steel in the ladle, and issues temperature warning signs based on temperature differences within the region.

[0086] This application also proposes a ladle infrared monitoring image recognition device, in which the ladle is located between the converter and the infrared thermal imager. The infrared thermal imager is connected to the image processing system, the control system is connected to the image processing system, and a display is also connected to the image processing system.

[0087] The ladle infrared monitoring image recognition device also includes a memory, which stores a computer program. When the computer program is executed by a processor, the ladle infrared monitoring image recognition method described above can be executed.

[0088] The image processing system may be a completely new design, or it may utilize an existing processor to execute the various steps of the overall wiring method of this application. The types of existing processing units may include, but are not limited to, a central processing unit (CPU), a digital signal processing unit (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices.

[0089] This application also provides a computer-readable storage medium storing computer program instructions thereon, which, when executed by a processor, implement the above-described steel ladle infrared monitoring image recognition method.

[0090] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for recognizing infrared monitoring images of steel ladles based on a fully convolutional neural network, characterized in that, include: Acquire one or more infrared thermal images of the ladle; The infrared thermal imaging image of the steel ladle is used to extract features using a fully convolutional neural network (FCN), and outputs a first feature image corresponding to the low-resolution infrared thermal imaging image of the steel ladle, and / or a second feature image corresponding to the high-resolution infrared thermal imaging image of the steel ladle. Coarse segmentation is performed on the first feature image and / or the second feature image, wherein in the coarse segmentation process, one-dimensional feature values ​​in the first feature image and / or the second feature image are segmented and features are extracted. The first feature image and / or the second feature image after coarse segmentation are further segmented using Gibbs distribution. In the fine segmentation process, the first feature image and / or the second feature image after coarse segmentation are associated and identified and feature extracted by combining the context of the infrared thermal imaging image of the ladle. as well as The performance of the first feature image and / or the second feature image after fine segmentation is evaluated.

2. The method for recognizing infrared monitoring images of steel ladles based on a fully convolutional neural network according to claim 1, characterized in that, Also includes: Preprocessing is performed on the acquired infrared thermal imaging image of the ladle. The feature extraction is performed on the preprocessed infrared thermal imaging image of the ladle using a fully convolutional neural network (FCN). The preprocessing includes: for infrared thermal imaging images with blurred edges and / or blurred features, pre-processing edge pixel supplementation, image noise reduction and / or enhancement is performed.

3. The method for recognizing infrared monitoring images of steel ladles based on a fully convolutional neural network according to claim 1, in, Feature extraction from the infrared thermal imaging image of the ladle using a fully convolutional neural network (FCN) includes: The clarity of the infrared thermal imaging image of the ladle is detected; Based on the detection results of the aforementioned clarity, the infrared thermal imaging image of the ladle is identified as either a low-resolution infrared thermal imaging image or a high-resolution infrared thermal imaging image; and The low-resolution infrared thermal imaging image is subjected to upsampling / deconvolution processing using an upsampling block, and the high-resolution infrared thermal imaging image is subjected to downsampling / convolution processing using a downsampling block.

4. The method for recognizing infrared monitoring images of steel ladles based on a fully convolutional neural network according to claim 1, characterized in that, Also includes: In-depth learning is performed on the infrared thermal imaging images of the steel ladle that meet the performance evaluation criteria.

5. The method for recognizing infrared monitoring images of steel ladles based on a fully convolutional neural network according to claim 1, characterized in that, The performance of the infrared thermal imaging images of the finely segmented ladle is evaluated using at least one of the following: accuracy, recall, F-β index, and J index.

6. The method for recognizing infrared monitoring images of steel ladles based on a fully convolutional neural network according to claim 5, characterized in that, When evaluating the accuracy (P), recall, and F of infrared thermal imaging images of ladles obtained through fine segmentation processing... β When both the index and the J index are below the specified threshold, it is determined that the feature recognition of the infrared thermal imaging image of the ladle obtained through fine segmentation is poor, and the infrared thermal imaging image of the ladle needs to be iteratively identified again using the FCN algorithm through steps S12-S14.

7. A steel ladle infrared monitoring image recognition device, characterized in that, include: Memory and processor; The memory stores a computer program, which, when executed by the processor, performs the steel ladle infrared monitoring image recognition method according to any one of claims 1-6.

8. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the steel ladle infrared monitoring image recognition method according to any one of claims 1-6.

Images