Automatic identification method for flying steel of bar and wire in rolling mill, storage medium and program product

Abstract

The application discloses a rolling mill bar wire flying steel automatic identification method, a storage medium and a program product, relates to the technical field of image data processing, and comprises the following steps: collecting environment parameters of a rolling mill; acquiring a video image of the rolling mill which is preprocessed by a polarizing light sheet; constructing a detection model; inputting the video image into a feature extraction layer of the detection model to extract local features of the flying steel and output spatial features; inputting a continuous video frame feature sequence into a time sequence modeling layer of the detection model to output time sequence features; and a decision fusion layer of the detection model dynamically adjusts weights based on the environment parameters to fuse a CNN standardized feature vector and a Transformer standardized feature vector to obtain a fusion feature vector and make a classification decision to output a flying steel detection result, wherein the flying steel detection result comprises a probability and position coordinates of the flying steel. The application solves the problems of missed detection and false detection of the flying steel caused by image blur in traditional machine vision detection of the flying steel.

Description

Technical Field

[0001] This invention relates to the field of image data processing technology, specifically to an automatic identification method, storage medium, and program product for flying steel bars and wire rods in a steel rolling mill. Background Technology

[0002] High-speed bar and wire rod generally refers to bars and wire rods rolled at a speed of ≥30 m / s. The production of high-speed bar and wire rod is a complex process that involves processing steel billets into high-precision, high-speed wire rods through multiple rolling processes. The main process of high-speed bar and wire rod production is as follows: the billet is conveyed to the walking beam furnace via roller conveyor, and the heating temperature is controlled at 1150~1250℃ (adjusted according to the steel grade) to ensure uniform austenitization of the billet and reduce rolling cracks; then it is rolled through roughing (composed of 6~8 rolling mills (such as alternating horizontal and vertical arrangement), rolling the billet from square billet into round billet with a diameter of 50~60mm, with a rolling speed of about 5~15 m / s), intermediate rolling (4~6 rolling mills, further rolling into intermediate billet with a diameter of 20~30mm, with the speed increased to 20~40 m / s), and finishing rolling (usually composed of 10~12 rolling mills with a top angle of 45° and no twist (such as Morgan type rolling mill), rolling the rolled piece into the finished size (Φ5~Φ20mm), with a rolling speed of up to 80~120 m / s). After controlled rolling and controlled cooling (TMCP), it is finally finished and shipped out.

[0003] Currently, during the rolling process of bar and wire rod production lines, there is a phenomenon where rolled pieces detach from the roller table due to abnormal stress, ejecting from the rolling mill area at high speeds (up to tens of meters per second). This flying steel not only causes equipment damage and production line shutdowns, but can also lead to safety accidents such as burns and impacts, posing a significant risk to bar and wire rod production.

[0004] Due to the complex working conditions in the steel rolling workshop, such as high temperature, water vapor, and strong light reflection, traditional machine vision detection of flying steel is prone to problems of missed detection and false detection because the images are blurry. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, a method, storage medium, and program product for automatic identification of flying steel in bar and wire rods in steel rolling workshops are provided to solve the problem that traditional machine vision detection of flying steel is prone to missed detection and false detection due to image blurring.

[0006] To achieve the above objectives, an automatic identification method for fly-through steel bars and wire rods in a rolling mill is provided, comprising:

[0007] Environmental parameters of the steel rolling workshop are collected, including relative humidity and light intensity.

[0008] Acquire video images of the steel rolling workshop that have undergone polarization preprocessing;

[0009] A detection model is constructed, which includes a feature extraction layer, a temporal modeling layer, and a decision fusion layer.

[0010] The video image is input into the feature extraction layer, which uses CNN to extract local features of the flying steel to output spatial features;

[0011] Based on the rolling speed, the local spatial feature maps of single frames of flying steel in multiple consecutive frames of the spatial features are spliced ​​together in time stamp order to form a continuous video frame feature sequence.

[0012] The continuous video frame feature sequence is input into the temporal modeling layer, which learns the long-distance dependency of the flying steel motion through the Transformer encoder to output temporal features;

[0013] The decision fusion layer standardizes the spatial features and the temporal features to obtain CNN-standardized feature vectors and Transformer-standardized feature vectors;

[0014] Based on the environmental parameters, the decision fusion layer dynamically adjusts the weights to fuse the CNN-normalized feature vector and the Transformer-normalized feature vector according to the adjusted weights to obtain a fused feature vector;

[0015] The decision fusion layer inputs the fused feature vector into the multilayer perceptron for classification decision-making to output the flying steel detection result, which includes the probability of flying steel occurrence and location coordinates.

[0016] Furthermore, the video images are captured by an industrial camera, and an optical polarizer is mounted in front of the lens of the industrial camera.

[0017] Furthermore, the CNN is a lightweight CNN.

[0018] Furthermore, the step of dynamically adjusting the weights includes:

[0019] Based on the environmental parameters, the degree of environmental interference is determined, including strong interference, moderate interference, and weak interference.

[0020] When the environmental interference level is determined to be strong interference, the decision fusion layer adjusts the weights of the CNN standardized feature vector to 0.1~0.4 and the weights of the Transformer standardized feature vector to 0.6~0.9.

[0021] When the level of environmental interference is determined to be weak interference, the decision fusion layer adjusts the weights of the CNN standardized feature vector to 0.6~0.7 and the weights of the Transformer standardized feature vector to 0.3~0.4.

[0022] When the level of environmental interference is determined to be moderate, the decision fusion layer dynamically adjusts the weight of the Transformer standardized feature vector by linear interpolation, increasing it linearly from 0.4 to 0.6.

[0023] The present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the aforementioned method.

[0024] The present invention provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the aforementioned method.

[0025] The beneficial effects of this invention are that the automatic identification method for flying steel in bar and wire rods in the rolling mill workshop can accurately identify high-speed flying steel under complex working conditions such as high temperature, water vapor, and strong light reflection in the rolling mill workshop. This is achieved by using a hardware fog-penetrating polarized light sheet and a CNN+Transformer fusion detection model. This method breaks through the limitations of traditional CNN models in capturing the temporal features of high-speed movement and irregular trajectories of flying steel. By utilizing the long-distance dependency modeling capability of the Transformer architecture, the method improves the global correlation analysis capability of flying steel splash trajectory and motion pattern.

[0026] The automatic identification method for flying steel bars and wire rods in the rolling mill of this invention utilizes industrial cameras in the rolling mill combined with polarizing filters to replace high-speed explosion-proof cameras with low-cost hardware modifications. At the same time, it leverages the efficient feature learning capabilities of Transformer to reduce the computational resource consumption caused by stacking multiple models, thereby meeting the real-time detection requirements of industrial edge devices.

[0027] The automatic identification method for fly-through steel in bar and wire rod production workshops of the present invention reduces equipment damage, downtime losses and safety accidents by detecting fly-through steel faults in real time and accurately, while simultaneously improving the integrity of production data, providing reliable support for process optimization and reducing raw material and energy costs. Attached Figure Description

[0028] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0029] Figure 1 This is a flowchart illustrating the automatic identification method for flying steel bars and wire rods in a steel rolling workshop according to an embodiment of the present invention.

[0030] Figure 2 This is an architecture diagram of the detection model according to an embodiment of the present invention.

[0031] Figure 3 This is a schematic diagram of the structure of a computer program product according to an embodiment of the present invention. Detailed Implementation

[0032] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0033] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0034] Reference Figure 1 and Figure 2 As shown, this invention provides an automatic identification method for fly-off steel in bar and wire rod production lines, comprising the following steps:

[0035] S1. Collect environmental parameters of the steel rolling workshop, including relative humidity and light intensity.

[0036] In this embodiment, environmental parameters are collected through a multi-source data acquisition system. Specifically, the multi-source data acquisition system integrates existing industrial cameras (frame rate ≥120fps), rolling force sensors, and temperature sensors in the existing rolling mill area, and simultaneously acquires video streams and process parameters (such as rolling speed and tension).

[0037] Relative humidity is collected by humidity sensors installed in the rolling mill, such as the humidity sensor outputting 60%~85% relative humidity in the finishing rolling area of ​​the rolling mill.

[0038] Light intensity can be measured by the strong light reflection intensity value output by the photosensitive sensor built into the industrial camera, measured in lux.

[0039] S2. Obtain video images of the steel rolling workshop that have been preprocessed with polarizers.

[0040] In this embodiment, video images are captured by an industrial camera. An optical polarizer is mounted in front of the lens of the industrial camera.

[0041] By installing an adjustable polarization angle optical polarizer in front of the industrial camera lens, the image clarity is improved by suppressing specular reflection light generated by high-temperature water vapor. In particular, it effectively improves the image blurring problem caused by cooling water mist in the finishing rolling zone (high-speed rolling scenario of 80~120 m / s).

[0042] In some embodiments, image preprocessing further includes grayscale and normalization of the video stream to reduce noise interference.

[0043] S3. Construct a detection model, which includes a feature extraction layer, a temporal modeling layer, and a decision fusion layer.

[0044] See Figure 2 As shown, in this embodiment, the detection model is a CNN+Transformer fusion detection model. The preprocessed image features are spatiotemporally aligned with real-time rolling force and temperature data, and then input into the CNN+Transformer model for multimodal feature fusion.

[0045] S4. Input the video image into the feature extraction layer. The feature extraction layer uses CNN (Convolutional Neural Network) to extract local features of the flying steel to output spatial features.

[0046] Among them, the CNN is a lightweight CNN. The CNN output is a "single-frame spatial feature map" (focusing on the noise ratio, such as the area of ​​blurred regions caused by water mist).

[0047] CNN extracts local features of the flying steel (such as edges and shape), and introduces residual connections and attention mechanisms to enhance detail capture in response to the blurring effect caused by the high-speed movement of the flying steel.

[0048] S5. Based on the rolling speed, the local spatial feature maps of single frames of flying steel in multiple consecutive frames of spatial features are spliced ​​together in time stamp order to form a continuous video frame feature sequence.

[0049] S6. Input the continuous video frame feature sequence into the temporal modeling layer. The temporal modeling layer learns the long-distance dependency of the flying steel motion through the Transformer encoder to output temporal features.

[0050] The Transformer encoder takes the feature sequence of consecutive video frames as input and learns the long-range dependencies of the flying steel motion (such as the trajectory association between adjacent frames) through a self-attention mechanism, thus solving the motion estimation error problem of the traditional optical flow method in high-speed scenes.

[0051] The "temporal feature vector" output by the Transformer encoder focuses on trajectory coherence, such as the deviation value of the flying steel position in adjacent frames.

[0052] S7, the decision fusion layer standardizes spatial and temporal features to obtain CNN-standardized feature vectors and Transformer-standardized feature vectors.

[0053] In some embodiments, the spatial feature vector output by the CNN (converted from the feature map to a 1-dimensional vector through global average pooling, such as 128-dimensional vector in the embodiment) and the temporal feature vector output by the Transformer (a 1-dimensional vector output by the last layer of the encoder, such as 512-dimensional vector in the embodiment) are L2 normalized to eliminate the influence of dimensionality differences.

[0054] S8. Based on environmental parameters, the decision fusion layer dynamically adjusts the weights to fuse the CNN-standardized feature vector and the Transformer-standardized feature vector according to the adjusted weights to obtain the fused feature vector.

[0055] In this embodiment, the step of dynamically adjusting the weights includes:

[0056] Based on environmental parameters, the degree of environmental interference is determined, which includes strong interference, moderate interference, and weak interference.

[0057] When the level of environmental interference is determined to be strong interference, the decision fusion layer adjusts the weights of the CNN standardized feature vectors to 0.1~0.4 and the weights of the Transformer standardized feature vectors to 0.6~0.9.

[0058] When the level of environmental interference is determined to be weak interference, the decision fusion layer adjusts the weights of the CNN standardized feature vectors to 0.6~0.7 and the weights of the Transformer standardized feature vectors to 0.3~0.4.

[0059] When the level of environmental interference is determined to be moderate, the decision fusion layer dynamically adjusts the weights of the Transformer standardized feature vectors linearly from 0.4 to 0.6 using linear interpolation.

[0060] Specifically, in some embodiments, the decision fusion layer first establishes an "environmental parameter-weight mapping relationship," including the following steps:

[0061] The collected environmental parameters are converted into "CNN weight coefficients (W_CNN)" and "Transformer weight coefficients (W_Trans)", and the core mapping rules are as follows:

[0062] When environmental interference is weak (e.g., humidity ≤60%, no strong light reflection, clear CNN spatial features, and noise ratio <5%), dynamically adjust W_CNN = 0.6~0.7 and W_Trans = 0.3~0.4. In this case, rely on the clear spatial features of CNN to dominate the decision-making, reducing the computational load of Transformer.

[0063] When environmental interference is strong (such as humidity > 75%, strong light reflection causing CNN features to be blurred, and noise accounting for ≥ 15%): dynamically adjust W_CNN = 0.1~0.4, W_Trans = 0.6~0.9. At this time, rely on the temporal features of Transformer (such as the continuity of the trajectory of flying steel) to make up for the lack of spatial features of CNN and reduce misjudgment caused by blurred images.

[0064] When the interference is moderate: the weights are dynamically adjusted by linear interpolation (e.g., when the humidity is 60%~75%, W_Trans increases linearly from 0.4 to 0.6 as the humidity increases).

[0065] In some embodiments, the dynamic weighted calculation of the decision fusion layer includes the following steps:

[0066] Calculate the fused feature vector F_fusion = W_CNN × F_CNN + W_Trans × F_Trans.

[0067] Where F_CNN is the CNN-normalized feature vector;

[0068] F_Trans is the Transformer standardized feature vector, and W_CNN+W_Trans=1.

[0069] S9, the decision fusion layer, inputs the fused feature vector into the multilayer perceptron for classification decision-making to output the flying steel detection result, which includes the probability of flying steel occurrence and the location coordinates.

[0070] The decision fusion layer integrates the spatial features extracted by CNN and the temporal features output by Transformer, and performs classification decisions through a multilayer perceptron (MLP), outputting the probability of flying steel occurrence (0~1) and location coordinates, such as the pixel coordinates of the feature map corresponding to the original image.

[0071] The model outputs the confidence score of the flying steel, and performs dual verification by combining a preset threshold (such as confidence score > 0.8) and abnormal process parameters (such as a sudden drop in rolling force) to reduce the false judgment rate.

[0072] When the model determines that a steel flying incident has occurred, an alarm signal is triggered in real time, the flying shear action is controlled, and the steel flying incident video and process data are simultaneously recorded to the server to execute an emergency response.

[0073] To avoid the limitations of fixed mapping rules, the mechanism also includes real-time feedback optimization logic:

[0074] When the model outputs the detection results of flying steel that are consistent with the manually labeled results (e.g., no false positives / false negatives for 100 consecutive frames), the current "environmental parameter-weight mapping relationship" is maintained.

[0075] When the manually labeled flying steel samples are inconsistent with the model detection results, the model is automatically triggered to incrementally train and continuously optimize its ability to identify new flying steel forms (such as scrap broken by flying shears and disordered red high-temperature scrap steel in the hopper of overhead cranes).

[0076] Specifically, when misjudgment / missed judgment occurs (such as missed judgment due to insufficient Transformer weights in a high water vapor environment), the environmental parameters, weight values ​​and error type at this time are automatically recorded. The mapping parameters of the "interference evaluation model" are optimized through incremental learning, so that the weight allocation in the same environment is more accurate in the future (such as fine-tuning the lower limit of W_Trans in a high water vapor environment from 0.6 to 0.65).

[0077] The automatic identification method for flying steel in bar and wire rod in the rolling mill workshop of the present invention avoids the blind spots of traditional models in water mist environment, realizes global correlation analysis of flying steel dynamic trajectory, and improves the overall detection accuracy to 98.7%.

[0078] The automatic identification method for flying steel bars and wire rods in the rolling mill of this invention combines hardware defogging with Transformer time-series modeling to reduce the impact of environmental noise on detection. In high humidity environments, the false positive rate is reduced from 8.5% to 2.1%, and the missed detection rate in strong light reflection scenarios is reduced by 76%.

[0079] Compared to multi-model parallel schemes, the automatic identification method for fly-through steel bars and wire rods in the rolling mill of this invention reduces computational complexity by more than 30%, making it suitable for the real-time requirements of industrial edge devices. The detection latency of the automatic identification method for fly-through steel bars and wire rods in the rolling mill of this invention is ≤50ms, meeting the real-time response requirements of high-speed rolling scenarios at 120 m / s.

[0080] The automatic identification method for flying steel bars and wire rods in the rolling mill workshop of the present invention utilizes existing industrial cameras + polarizing filters to replace the high-cost solution of high-speed explosion-proof cameras, reducing hardware modification costs by 60% and reducing the deployment cost per workshop from 300,000 yuan to 120,000 yuan.

[0081] The automatic identification method for fly-through steel bars and wire rods in the rolling mill of this invention reduces downtime and equipment maintenance costs caused by fly-through steel bars, improves data integrity, and increases the data integrity rate of production quality traceability in the rolling mill from 65% to 92%.

[0082] The automatic identification method for flying steel bars and wire rods in the rolling mill of this invention solves the image blurring problem in high-temperature water mist environments by using a polarized light sheet hardware defogging technology. It enhances the ability to capture the temporal features of high-speed flying steel by leveraging a CNN+Transformer fusion model, breaking through the limitations of traditional models that rely on YOLO series or CNN architectures and lack long-distance dependency modeling. Furthermore, through a dynamic weight adaptive mechanism and incremental learning optimization module, it reduces computational complexity and hardware modification costs, solving the problems of low efficiency and high hardware deployment costs in multi-model parallel processing. It achieves comprehensive optimization in detection accuracy, anti-interference ability, computational efficiency, and cost control.

[0083] The present invention also provides a computer program product comprising computer instructions stored in a computer-readable storage medium and adapted to be read and executed by a processor to cause a computer device having the processor to perform the methods provided in any of the above embodiments.

[0084] Please see Figure 3 , Figure 3 The hardware structure of a computer program product according to an embodiment of the present invention is illustrated. The computer program product includes: a processor 1, a memory 2, an input / output interface 3, a communication interface 4, and a bus 5.

[0085] Processor 1 is implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present invention.

[0086] The memory 2 is implemented in the form of read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 2 stores the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 2 and is called by the processor 1 to execute the automatic charging method for betting stations executed by the control device in the embodiments of this invention, or to execute the automatic charging method for betting stations executed by the cloud server in the embodiments of this invention.

[0087] Input / output interface 3 is used to implement information input and output.

[0088] Communication interface 4 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0089] Bus 5 transmits information between various components of the device (e.g., processor 1, memory 2, input / output interface 3, and communication interface 4).

[0090] The processor 1, memory 2, input / output interface 3, and communication interface 4 communicate with each other within the device via bus 5.

[0091] The present invention provides a computer-readable storage medium storing at least one computer program instruction, which is loaded and executed by a processor to perform the operations performed by the method provided in any of the above embodiments.

[0092] The functions described herein can be implemented in hardware, software executed by a processor, firmware, or any combination thereof. If implemented in software executed by a processor, the functions can be stored as one or more instructions or codes on or transmitted via a computer-readable medium. Other examples and embodiments are within the scope and spirit of this invention and the appended claims. For example, due to the nature of software, the functions described above can be implemented using software executed by a processor, hardware, firmware, hardwired, or any combination thereof. Furthermore, the functional units can be integrated into a single processing unit, or each unit can exist physically separately, or two or more units can be integrated into a single unit.

[0093] In the embodiments provided by this invention, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection can be through some interfaces; the indirect coupling or communication connection of units or modules can be electrical or other forms.

[0094] The units described as separate components may or may not be physically separate. Similarly, the components of the control device may or may not be physical units; they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0095] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing computer program instructions, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0096] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A method for automatic identification of flying steel bars and wire rods in a steel rolling mill, characterized in that, include: Environmental parameters of the steel rolling workshop are collected, including relative humidity and light intensity. Acquire video images of the steel rolling workshop that have undergone polarization preprocessing; A detection model is constructed, which includes a feature extraction layer, a temporal modeling layer, and a decision fusion layer. The video image is input into the feature extraction layer, which uses a CNN to extract local features of the flying steel to output spatial features; Based on the rolling speed, the local spatial feature maps of single frames of flying steel in multiple consecutive frames of the spatial features are spliced ​​together in time stamp order to form a continuous video frame feature sequence. The continuous video frame feature sequence is input into the temporal modeling layer, which learns the long-distance dependency of the flying steel motion through the Transformer encoder to output temporal features; The decision fusion layer standardizes the spatial features and the temporal features to obtain CNN-standardized feature vectors and Transformer-standardized feature vectors; Based on the environmental parameters, the decision fusion layer dynamically adjusts the weights to fuse the CNN-normalized feature vector and the Transformer-normalized feature vector according to the adjusted weights to obtain a fused feature vector; The decision fusion layer inputs the fused feature vector into the multilayer perceptron for classification decision-making to output the flying steel detection result, which includes the probability of flying steel occurrence and location coordinates.

2. The automatic identification method for fly-off steel bars and wire rods in a rolling mill according to claim 1, characterized in that, The video images are captured by an industrial camera, and an optical polarizer is installed in front of the lens of the industrial camera.

3. The automatic identification method for fly-off steel bars and wire rods in a rolling mill according to claim 1, characterized in that, The CNN mentioned is a lightweight CNN.

4. The automatic identification method for fly-off steel bars and wire rods in a rolling mill according to claim 1, characterized in that, The steps for dynamically adjusting the weights include: Based on the environmental parameters, the degree of environmental interference is determined, including strong interference, moderate interference, and weak interference. When the environmental interference level is determined to be strong interference, the decision fusion layer adjusts the weights of the CNN standardized feature vector to 0.1~0.4 and the weights of the Transformer standardized feature vector to 0.6~0.

9. When the level of environmental interference is determined to be weak interference, the decision fusion layer adjusts the weights of the CNN standardized feature vector to 0.6~0.7 and the weights of the Transformer standardized feature vector to 0.3~0.

4. When the level of environmental interference is determined to be moderate, the decision fusion layer dynamically adjusts the weight of the Transformer standardized feature vector by linear interpolation, increasing it linearly from 0.4 to 0.

6.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method described in any one of claims 1 to 3.

6. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method described in any one of claims 1 to 3.

Images