A multi-receptive field-based lightweight strip steel surface defect prediction method

Abstract

The application discloses a strip steel surface defect detection method based on a convolutional neural network, and comprises the following steps: step 1, constructing an LSDD model, wherein the LSDD model comprises an encoder and a decoder, the encoder is composed of five MHL modules, and the decoder is composed of five LEF modules; step 2, training the LSDD model by using a public data set; and step 3, taking image data as input and performing prediction through the trained LSDD model. According to the method, the defect information of the strip steel surface can be restored to a greater extent on the premise of maintaining light weight (few parameters and fast operation rate).

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, specifically to a method for predicting surface defects in lightweight strip steel based on multiple receptive fields. Background Technology

[0002] Deep neural networks have achieved remarkable success in computer vision tasks such as image classification, object detection, and target tracking. However, with the advancement of technology, there is increasing focus on the performance of deep neural networks in practical applications. A clear trend in the evolution of artificial intelligence is the deployment of high-performance neural network models on edge platforms and their real-time execution (>30 frames per second) in real-world scenarios, especially on mobile and embedded devices. These platforms are characterized by limited memory resources, low processor performance, and power consumption constraints. Therefore, the most accurate models currently available are often difficult to deploy and run in real-time on these platforms. Due to limitations in storage space and computing power, the storage and computation of neural network models on mobile and embedded devices still face significant challenges. The paradox in this direction lies in how to balance performance with limited resources, i.e., maximizing performance under resource constraints.

[0003] To address this issue, lightweight networks have emerged, aiming to further reduce the number of model parameters and complexity while maintaining model accuracy, and have gradually become a research hotspot in the field of computer vision. Lightweight networks not only involve novel explorations of network structures but also utilize various model compression techniques, such as knowledge distillation and pruning, promoting the application of deep learning technology on mobile and embedded devices. Lightweight network models have made significant contributions in fields such as smart homes, security, autonomous driving, and video surveillance. The rapid deployment of lightweight networks on mobile devices has become an urgent problem to be solved in fields such as autonomous driving, robot vision, drones, facial recognition, and object detection.

[0004] Currently, most object detection architectures rely on convolutional networks for feature extraction, i.e., the backbone, such as the 8 layers of AlexNet, the 19 layers of VGGNet, the 22 layers of GoogleNet, and the 152 layers of ResNet. Although model performance improves with increasing network depth, these networks typically have enormous computational costs, especially on mobile hardware platforms with limited computing power (such as ARM, FPGA, RISC-V, and ASIC). Detection algorithms relying on these backbone networks struggle to meet real-time requirements. This leads to the frustrating situation where mobile devices cannot perform well, wearable devices cannot perform well, and data centers cannot afford to perform well. Therefore, accelerating object detection algorithms to meet the requirements of industrial applications remains a critical issue.

[0005] Traditional SOD methods require a powerful backbone (i.e., encoder) to capture low-level fine-grained details and high-level semantic features, and a carefully calibrated decoder to recover spatial resolution without losing spatial information. Both of these can lead to enormous parameter and computational overhead. Conversely, the recent increase in interest in mobile applications, such as smartphones, where computing power, memory, and energy support are limited, makes it impossible to deploy these large SOD models. Deploying these large models on servers is also very costly. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention proposes a method for predicting surface defects of lightweight strip steel based on multiple receptive fields, which restores the surface curve of strip steel as much as possible while achieving lightweight design.

[0007] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0008] A method for predicting surface defects in lightweight strip steel based on multiple receptive fields includes the following steps:

[0009] Step 1: Construct an LSDD model, which includes an encoder and a decoder. The encoder consists of 5 MHL modules, divided into 5 layers from top to bottom, denoted as encoder1 to encoder5, with features propagating from top to bottom. The decoder consists of 5 LEF modules, divided into 5 layers from top to bottom, denoted as decoder1 to decoder5, with input features propagating from bottom to top. Each LEF module is connected to a convolutional layer for channel compression.

[0010] Step 2: Train the LSDD model using a public dataset;

[0011] Step 3: Using image data as input, perform predictions through the trained LSDD model:

[0012] Step 3-1: Using image data as input, encoder1 outputs e1, which is then used as input for the next encoder layer. After passing through 5 MHL modules, multi-level features are obtained. That is, the output of the i-th encoder layer is denoted as e. i ;

[0013] Step 3-2: The output of each MHL module is passed to the corresponding decoder, i.e., the output e of the i-th decoder layer. i This will be passed as a parameter to the i-th encoder layer;

[0014] Step 3-3: Perform a 1×1 convolution on the outputs d1 to d5 of each LEF module to compress the channels to 1, and obtain the decoder's prediction maps out1 to out5;

[0015] Steps 3-4: Upsample the prediction images out1 to out5 of each level to the pixel size of the input image data, and then perform a 1×1 convolution. The input has 5 channels and the output has 1 channel, which gives the final prediction result of the model.

[0016] Preferably, the image data input in step 3 is 224*224 in size.

[0017] Preferably, the output is via the MHL module. i Method:

[0018] The input of the MHL module is fed in parallel into a convolutional combination of four receptive fields. The outputs of these four convolutional combinations are concat-fused. Then, the fused information is passed through a 1×1 convolution to compress the number of channels back to the input channel number, and this compressed information is added to the input features. Finally, a spatial attention module and a standard 3×3 convolution are used to output e. i .

[0019] Preferably, the convolutional combinations of the four receptive fields are as follows:

[0020] The first convolution combination is: a regular 1×1 convolution;

[0021] The second convolution combination is: a 1×1 depthwise separable convolution, a regular 1×1 convolution, a 3×3 depthwise separable convolution with a dilation of 3, and a regular 1×1 convolution.

[0022] The third convolution combination is: a 5×5 depthwise separable convolution, a regular 1×1 convolution, a 3×3 depthwise separable convolution with an inflation rate of 5, and a regular 1×1 convolution.

[0023] The fourth convolution combination is: a 7×7 depthwise separable convolution, a regular 1×1 convolution, a 3×3 depthwise separable convolution with a dilation of 7, and a regular 1×1 convolution.

[0024] Preferably, the LEF module has three inputs, namely the output of the MHL module. i The output d of the upper-layer LEF module i+1 And the prediction charts from the previous levels.

[0025] Preferably, the processing method of the LEF module is as follows:

[0026] First, output d from the upper-layer LEF module i+1 Perform a doubling upsampling so that the feature size is equal to e. i Keep them consistent and add them together. The result of the sum is compared with the prediction graph of the previous level {out5,out4,.....,out i-2 After upsampling to the same size, concatenate; then, pass each feature through a 1×1 convolution to scale the channels to the number of input channels, i.e., e. i The number of channels; after passing through a 3×3 depthwise separable convolution and a 1×1 convolution, the encoded and decoded information is fused to obtain the fused features;

[0027] Next, the fused features are fed into two parallel convolutional layers: one is a regular 3×3 convolution with half the original number of output channels; the other is a 3×3 convolution with a dilation rate of 2, also with half the original number of output channels. The outputs of the two parallel convolutional layers are then concatenated.

[0028] Finally, it goes through a regular 1×1 convolution, a 3×3 depthwise separable convolution, and another regular 1×1 convolution, with a spatial attention module added.

[0029] Preferably, in step 2, the training set of the public dataset SD-saliency-900 is used for training.

[0030] Preferably, the training method for the LSDD model is as follows: the LSDD model is deployed on an NVIDIA 2080ti server, the parameters are initialized with default settings during training, the optimizer used is Adam, and the learning rate is set to a fixed value of 0.00001.

[0031] Preferably, in step 2, the loss function used during training is the BCE loss, which is widely used for binary task classification and is defined as:

[0032]

[0033] Where G is the manually annotated supervised image, S is the predicted saliency map, and x and y are the horizontal and vertical coordinates of each pixel, respectively.

[0034] This invention has the following characteristics and beneficial effects:

[0035] The above technical solution is adopted:

[0036] While maintaining a lightweight design (fewer parameters, faster computation speed), this approach aims to reproduce defect information on the surface of strip steel to a greater extent. Specifically, a multi-scale hybrid receptive field lightweight coding block (MHL module) serves as the backbone network of the encoding stage, acquiring multi-scale information. The entire module uses only one complete 3×3 convolution, with the remaining operations employing low-parameter, high-efficiency methods such as depthwise separable convolution and 1×1 convolution. Depthwise separable convolution significantly reduces the number of parameters by decomposing the convolution operation into depthwise convolution and pointwise convolution. This makes the lightweight model more compact and suitable for operation in resource-constrained environments, such as mobile devices, embedded systems, and edge devices. Due to its structure, depthwise separable convolution has lower computational complexity compared to traditional convolution. It offers faster inference speed and lower power consumption, which is particularly important for mobile devices, embedded systems, or real-time image and speech recognition. On the decoder, the encoder layer captures high-level abstract features of the input. Summing these features helps integrate global and local information at different scales, improving the network's understanding of the input. The encoding process incorporates prediction maps from previous levels, enabling better acquisition of defect location information. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a diagram of the overall framework of the LSDD model in an embodiment of the present invention.

[0039] Figure 2 for Figure 1 MHL module framework diagram.

[0040] Figure 3 for Figure 1 LEF module framework diagram.

[0041] Figure 4 Examples demonstrating the effectiveness of the invention. Detailed Implementation

[0042] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0043] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0044] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0045] This embodiment provides a method for predicting surface defects in lightweight strip steel based on multiple receptive fields, such as... Figure 1 As shown, the model adopts an encoder-decoder U-shaped structure, specifically divided into two main parts: an encoder and a decoder. The final output is formed by fusing the predicted images output by the decoders at each level through convolutional layers. Specifically, the input image is first resized to 224×224 pixels and then fed into the encoder to obtain multi-level features. The encoder consists of five proposed MHL modules (Multi-Scale Hybrid Receptive Field Lightweight Encoding Blocks), divided into five layers from top to bottom, denoted as encoder1 to encoder5. After passing through these five encoder layers, the multi-level features are obtained. That is, the output of the i-th encoder layer is denoted as e. i For example, the output of encoder1 corresponds to e1. ​​It's important to note that the output of each layer is passed to the corresponding decoder layer, i.e., the output e of the i-th decoder layer. i This will be passed as a parameter to the i-th encoder layer to guide the decoder's gradual reconstruction.

[0046] The decoder consists of five proposed MHL (Multi-Scale Hybrid Receptive Field Lightweight Coding Blocks). The decoder employs the same hierarchical structure as the encoder, divided into five layers from top to bottom, denoted as decoder1 to decoder5, with the output of each layer labeled d1 to d5. In contrast to the encoder, the input features propagate from bottom to top, that is, from decoder5 to decoder1.

[0047] Specifically, a 1×1 convolution is performed on the outputs d1 to d5 of each encoder to compress the channels to 1, thus obtaining the prediction maps out1 to out5 of each decoder layer.

[0048] Finally, the predicted images out1 to out5 of each level are upsampled to the pixel size of the input image. Then, they are passed through a 1×1 convolution, with five input channels and one output channel, to obtain the final model output, which is the saliency map.

[0049] Traditional encoders such as VGG16 and ResNet have excessively large parameter counts and low computational efficiency. While lightweight coding modules from other domains, such as CSNet, are used, they fail to effectively extract semantic information from the complex defects on the surface of strip steel. It has been observed that the defects on the strip steel surface vary greatly in size and are unevenly distributed. To address these issues, a multi-scale hybrid receptive field lightweight coding block (MHL module) is proposed as the backbone network of the coding stage to acquire multi-scale information.

[0050] Specifically, it involves generating different levels of attention to objects of different scales, thereby actively capturing objects of similar size to the desired size. Therefore, the aim is to mimic the human eye's capture of objects of different scales by using different convolution sizes and dilation rates. Specifically, such as... Figure 2 As shown, the input of the MHL module is fed in parallel into four convolutional combinations with different receptive fields, as illustrated. Specifically, these are: a regular 1×1 convolution; a 1×1 depthwise separable convolution followed by a regular 1×1 convolution; a 3×3 depthwise separable convolution with a dilation of 3 followed by a regular 1×1 convolution; a 5×5 depthwise separable convolution followed by a regular 1×1 convolution; a 3×3 depthwise separable convolution with a dilation of 5 followed by a regular 1×1 convolution; and a 7×7 depthwise separable convolution followed by a regular 1×1 convolution; a 3×3 depthwise separable convolution with a dilation of 7 followed by a regular 1×1 convolution. Finally, the outputs of these four convolutional combinations are concatted and fused.

[0051] Using convolutions of different sizes and dilation rates allows for the acquisition of information from different receptive fields. The intermediate 1×1 convolutions are used to better fuse information from different channels. By using convolution kernels of varying sizes, multi-scale receptive fields can be achieved, enabling the model to capture features at different scales simultaneously. This is advantageous for handling targets or structures of different sizes, helping to improve the network's sensitivity to information at various scales in images and capture contextual information across different ranges, thus better fusing local and global features. This helps increase the effective receptive field of each convolution kernel, improving the network's understanding of different parts of the image and capturing the details and structure of the input data. In particular, using depthwise separable convolutions instead of ordinary convolutions aims to reduce the number of parameters and improve computational speed without sacrificing model accuracy as much as possible.

[0052] Next, the fused information is passed through a 1×1 convolution to compress the number of channels to the input channel count. Then, it is added to the input features. The use of a concat structure instead of simple addition aims to prevent the different scales of speech information from being encompassed by simple addition and subtraction. The subsequent 1×1 convolution aims to better fuse information from multiple scales, and compressing the channels to the input channel count also reduces the number of parameters. Adding to the input features follows the classic residual structure, which offers benefits such as solving the vanishing gradient problem, accelerating training convergence, and reducing the risk of overfitting.

[0053] Then, a spatial attention module and a regular 3×3 convolution are used. The spatial attention module is very lightweight, while extracting the most prominent semantic information in space, and can highlight the information obtained at different scales.

[0054] It can be observed that the entire module uses only a single complete 3×3 convolution, while the rest employ efficient methods with small parameter counts, such as depthwise separable convolutions and 1×1 convolutions. Depthwise separable convolutions significantly reduce the number of parameters by breaking down the convolution operation into depthwise and pointwise convolutions. This makes the lightweight model more compact and suitable for operation in resource-constrained environments, such as mobile devices, embedded systems, and edge devices. Due to its structure, depthwise separable convolutions have lower computational complexity than traditional convolutions. This results in faster inference speeds and lower power consumption, which is particularly important for mobile devices, embedded systems, and real-time image and speech recognition.

[0055] Experiments show that the MHL module can maintain high accuracy with a relatively small number of parameters.

[0056] Traditional encoding and decoding models mostly involve simply adding the encoded and decoded information together and then fusing them using simple convolution.

[0057] In this embodiment, the LEF module has three inputs, such as... Figure 3 As shown, these are the encoder layer inputs e. i Decoder layer input d i+1 The prediction maps of the previous layers correspond to the output of the encoder at layer i, the output of the decoder at layer (i+1), and the prediction maps of the previous layers, respectively.

[0058] First, the output d of the decoder layer... i Performing a doubling upsampling restores spatial details, making the feature size equal to the encoder layer input e. i To maintain consistency, match it with the output e of the coding layer. i Adding them together. The encoder layer captures high-level abstract features of the input. Adding them together helps to integrate global and local information at different scales, improving the network's understanding of the input.

[0059] A key characteristic of surface defects in strip steel is their highly disordered distribution. To better obtain defect location information, prediction maps from previous layers are incorporated into the encoding process. Specifically, the result obtained by adding the above encoding and decoding layers is compared with the prediction maps from previous layers {out5,out4,.....,out i-2 After upsampling to the same size, concatenate. It's worth noting that the previous level of prediction maps refers to the maps starting from out5 and moving upwards until out5. i+2 This refers to the prediction graph of the decoder at layer (i+2). The reason for removing the prediction graphs of adjacent layers is... i+1 This is because the prediction information from the previous level has already been output by the decoding layer. i+1 This information is provided instead of being provided by the prediction graph. Guided by the coarse predictions of previous layers, even shallower encoders can be guided by low-level, high-level semantic information. The reason for not directly using the output d from previous layers is... i The guidance is implemented for two reasons: first, directly introducing features would involve a large number of parameters because deeper feature values ​​have more channels; second, the coarse prediction map is already able to provide the locations of significant targets extracted from deeper layers to shallower layers.

[0060] The features are then sequentially processed by a 1×1 convolution to scale the channels to the number of input channels, i.e., e. i The number of channels is then increased. After that, a 3×3 depthwise separable convolution and a 1×1 convolution are performed to fuse the encoded and decoded information, which helps to further integrate feature information.

[0061] Next, again to capture multi-scale information, the feature values ​​are fed into two parallel convolutional layers: one is a standard 3×3 convolution with half the original number of output channels; the other is a 3×3 convolution with a dilation rate of 2, also with half the original number of output channels. Finally, the outputs of the two parallel convolutional layers are concatenated. The reason for halving the number of channels is to reduce the number of parameters in the bulky 3×3 convolution by half. The different dilation rates are used to expand the receptive field, capturing information at different scales and helping to capture a wider range of contextual information.

[0062] Finally, the data is processed sequentially through a regular 1×1 convolution, a 3×3 depthwise separable convolution, and another regular 1×1 convolution, with the aim of fusing information from different receptive fields.

[0063] Finally, similar to the encoding stage, a spatial attention module was added. This module is lightweight enough, and the spatial attention mechanism helps the model pay more attention to important regions in the input image, improving the model's perception of the local structure of the image and highlighting the details of the image to a greater extent.

[0064]

[0065] Table 1: Comparison of LSDD Experimental Results

[0066] As shown in Table 1, LSDD has achieved a good balance between overall performance, model parameter size, and computational efficiency in recent years. LSDD performs excellently on several key metrics, demonstrating its superior performance in salient object detection. It boasts high accuracy, robustness, and speed. LSDD has a relatively low parameter size of only 6.53M, making it a relatively lightweight model compared to other methods. The lower parameter size helps reduce the model's storage requirements and memory footprint, making it particularly suitable for resource-constrained environments such as mobile and edge devices. LSDD's inference speed (140 FPS) is relatively high, meaning it has good responsiveness in real-time applications. This high speed makes LSDD more suitable for tasks requiring fast response times, such as real-time video analytics and mobile applications. Despite being a lightweight model, LSDD excels on several performance metrics. Low MAE values ​​and high WF, OR, SM, and PFOM metrics indicate that LSDD achieves excellent performance in salient object detection tasks. This demonstrates that LSDD maintains a high level of accuracy and robustness while maintaining a lightweight design.

[0067] This embodiment shows a real image of a strip steel surface, the output results of various models, and the results of manual annotation, such as... Figure 4As shown, specifically from left to right: BASNet, CSNet, DACNet, EDRNet, Fast-SCNN, GT (human-annotated), HVPNet, LSDD, and the input image. It can be seen that our model, while maintaining a lightweight design, can largely reproduce surface details, achieving results comparable to or even better than ordinary models (BASNet, CSNet, DACNet, EDRNet). Furthermore, it significantly surpasses other lightweight models (Fast-SCNN, HVPNet), both in terms of the overall location of defect annotations and the refinement of edges.

[0068] In summary, LSDD has achieved success in lightweight design, striking a balance between parameter count, computational speed, and performance. This makes the LSDD model a preferred choice for resource-constrained scenarios such as mobile devices and edge computing.

Claims

1. A method for predicting surface defects in lightweight strip steel based on multiple receptive fields, characterized in that, Includes the following steps: Step 1: Construct an LSDD model, which includes an encoder and a decoder. The encoder consists of 5 MHL modules, arranged in five layers from top to bottom, denoted as... Features propagate from top to bottom; the decoder consists of 5 LEF modules, divided into 5 layers from top to bottom, denoted as follows: The input features are propagated from bottom to top; each LEF module is connected to a convolution for channel compression; The processing method of the LEF module is as follows: First, output from the upper-layer LEF module Perform upsampling by a factor of two, so that the feature size is equal to... Maintain consistency and add them together. The result of the addition is compared with the prediction graph of the previous level. After upsampling to the same size, perform concat. Then the features pass through one in sequence. Convolution scales the number of channels to the number of channels in the input. The number of channels; after one Depth-separable convolution, a Convolution is used to fuse encoded and decoded information to obtain fused features. Next, the fused features are fed into two parallel convolutional layers: one is a regular convolutional layer. One convolution has half the number of output channels; the other has a dilation rate of 2. The convolution of the two parallel convolutional layers results in half the number of output channels. Finally, it passes through a regular... Convolution, one Depthwise separable convolution, a regular Convolution, with the addition of a spatial attention module; Step 2: Train the LSDD model using a public dataset; Step 3: Using image data as input, perform predictions through the trained LSDD model: Step 3-1: Take the image data as input, and then... Output This data is then used as input to the next encoder layer, and after passing through 5 MHL modules, multi-level features are obtained. That is, the output of the i-th encoder layer is denoted as ; The input of the MHL module is fed in parallel into a convolutional combination of four receptive fields. The outputs of the four convolutional combinations are then concatenated and fused. Finally, the fused information is processed through a... The convolutional layer compresses the number of channels to the input channel number and adds it to the input features. Finally, a spatial attention module and a regular... Convolution output ; Step 3-2: The output of each MHL module is passed to the corresponding decoder, i.e., the output of the i-th decoder layer. This will be passed as a parameter to the i-th encoder layer; Step 3-3, Output of each LEF module Perform a The convolution process compresses the channels to 1, resulting in the decoder's prediction map. ; Steps 3-4: Create prediction maps for each level. Upsampled to the pixel size of the input image data, and then processed by a... The convolution, with 5 channels as input and 1 channel as output, yields the final prediction result from the model.

2. The method for predicting surface defects of lightweight strip steel based on multiple receptive fields according to claim 1, characterized in that, The image data input in step 3 is 224*224 in size.

3. The method for predicting surface defects of lightweight strip steel based on multiple receptive fields according to claim 2, characterized in that, The four convolutional combinations of the receptive fields are as follows: The first convolution combination is: a regular Convolution; The second convolution combination is: one Depthwise separable convolution, ordinary Convolution, a Depth-separable convolution with an inflation rate of 3, ordinary Convolution; The third convolution combination is: one Depthwise separable convolution, ordinary Convolution, a The dilation rate is 5, and the depthwise separable convolution is ordinary. Convolution; The fourth convolution combination is: one Depthwise separable convolution, ordinary Convolution, a The dilation rate is 7 for depthwise separable convolution, and ordinary... The convolution.

4. The method for predicting surface defects of lightweight strip steel based on multiple receptive fields according to claim 1, characterized in that, The LEF module has three inputs, namely the output of the MHL module. Output of the upper-layer LEF module And the prediction charts from the previous levels.

5. The method for predicting surface defects of lightweight strip steel based on multiple receptive fields according to claim 1, characterized in that, The training method for the LSDD model is as follows: the LSDD model is deployed on an NVIDIA 2080ti server, the parameters are initialized with default settings during training, the optimizer used is Adam, and the learning rate is set to a fixed value of 0.00001.

6. The method for predicting surface defects of lightweight strip steel based on multiple receptive fields according to claim 5, characterized in that, In step 2, the loss function used during training is BCE loss, which is widely used for binary task classification and is defined as: ; Where G is the manually labeled supervised image, S is the predicted saliency map, and x and y represent the horizontal and vertical coordinates of each pixel, respectively.

Images