Lightweight reduction furnace atomization detection method and device based on target detection technology, electronic equipment and storage medium

Abstract

The invention relates to the technical field of polycrystalline silicon preparation, in particular to a lightweight reduction furnace atomization detection method and device based on a target detection technology, electronic equipment and a storage medium, and the method comprises the following steps: obtaining an in-furnace image of a reduction furnace, and preprocessing the in-furnace image of the reduction furnace to obtain the preprocessed in-furnace image of the reduction furnace; a reduction furnace atomization detection model is constructed and trained, and a trained reduction furnace atomization detection model is obtained; and inputting the preprocessed operation image of the reduction furnace into the trained atomization detection model of the reduction furnace to obtain an atomization detection result. According to the method, the atomization detection precision is improved, meanwhile, the parameter quantity and complexity of the model are reduced, the detection speed is increased, and the requirement for the real-time performance of industrial detection is met.

Description

Technical Field

[0001] The present invention relates to the technical field of polysilicon preparation, and is a lightweight reduction furnace atomization detection method, device, electronic equipment and storage medium based on target detection technology. Background Art

[0002] With the continuous development of my country's photovoltaic industry, my country's polysilicon industry has shown a rapid development trend, and its output has continued to grow. The polysilicon reduction furnace is an important equipment for manufacturing polysilicon ingots. During the production process of the polysilicon reduction furnace, the change of the atomization state in the furnace has an important impact on product quality and production efficiency. The detection of the atomization state in the furnace is of great significance for optimizing the production process, improving product quality and production efficiency. At present, various companies mainly use manpower to conduct regular inspections at various reduction sites, but the reduction site has the characteristics of high noise, high pollution and certain danger. In addition, traditional target detection methods often rely on manual visual detection or traditional rule-based machine learning algorithms, which have limitations in processing large-scale data and complex atomization situations.

[0003] In recent years, fog detection models based on deep learning technology have made significant progress, but their computational complexity and running speed are still limited.

[0004] The introduction of lightweight neural network technology provides a new solution for reduction furnace atomization detection. It has a simple structure, fewer parameters, and fast running speed. It has good applicability on embedded devices with limited computing resources and can meet the needs of edge computing and real-time detection. Although lightweight neural networks have advantages on resource-constrained embedded devices, existing models often perform poorly when faced with complex atomization images, and are prone to false positives or missed detections. Summary of the invention

[0005] The present invention provides a lightweight reduction furnace atomization detection method, device, electronic device and storage medium based on target detection technology, which overcomes the shortcomings of the above-mentioned prior art and can effectively solve the problem that the existing models often perform poorly when facing complex atomization images and are prone to false alarms or missed detections.

[0006] One of the technical solutions of the present invention is achieved by the following measures: A lightweight reduction furnace atomization detection method based on target detection technology comprises the following steps:

[0007] Acquire the image inside the reduction furnace, preprocess the image inside the reduction furnace, and obtain the preprocessed image inside the reduction furnace;

[0008] Constructing and training a reduction furnace atomization detection model to obtain a trained reduction furnace atomization detection model;

[0009] The preprocessed reduction furnace operation image is input into the trained reduction furnace atomization detection model to obtain the atomization detection result.

[0010] The following is a further optimization and / or improvement of one of the above-mentioned technical solutions:

[0011] The above-mentioned reduction furnace atomization detection model may include an input end, a Backbone, a Neck, and a Head. The Backbone replaces the C2F module with a GS-C2F module based on the feature extraction network of the YOLOv8 network and the path aggregation network with a feature pyramid network, adds a CBAM module after the Backbone layer, and replaces the CIOU function in the loss function with a WIOU loss function.

[0012] The above-mentioned GS-C2F module may include a first CBS layer, a Spilt layer, a GhostBottleneck layer, a splicing layer, a second CBS layer and a SE layer connected in sequence. The input feature of the GS-C2F module passes through the first CBS layer to obtain the first feature. The Spilt layer divides the first feature into the second feature and the third feature in half according to the channel. The third feature passes through the GhostBottleneck layer to obtain the fourth feature. The second feature, the third feature and the fourth feature are dimensionally merged through the splicing layer, and finally calculated through the SE layer to obtain the output feature of the GS-C2F module.

[0013] The above GhostBottleneck layer integrates Ghost convolution, which can divide the input into two parts: ordinary convolution and simple linear operation for feature extraction, and finally perform concat operation to fuse the features to obtain the final output; Ghost convolution first extracts the information features between channels through convolution, and then uses grouped convolution to generate a new feature map; ordinary convolution is used to generate a feature map with a smaller channel, and then a linear transformation is performed; the formula for ordinary convolution is:

[0014] O=I*f

[0015] Among them, * represents the convolution operation, f is the convolution kernel, k*k is the convolution kernel size, O is the output feature map, and then L i,j Operate to produce Ghost feature map from each channel feature map of O;

[0016]

[0017] Among them, represents the i-th intrinsic feature map, represents the linear operation responsible for generating the j-th ghost map; finally, the feature maps obtained by ordinary convolution and linear transformation are spliced ​​together to get the final output.

[0018] The first CBS layer and the second CBS layer may both adopt a CBS structure, and the CBS structure includes a convolutional layer, a batch normalization layer, and a silu activation function layer connected in sequence.

[0019] The above-mentioned SE layer can learn the global context information of the input features of the GS-C2F module; the SE layer includes a Squeeze part and an Excitation part, wherein the Squeeze part uses a global average pooling operation to compress the spatial information of each feature map into a channel descriptor to reflect the overall activity level of the channel; wherein the Excitation part captures the dependencies between channels through two fully connected layers and learns the weights of each channel to enhance or suppress specific feature responses.

[0020] The above-mentioned preprocessing method may include adjusting the size of the reduction furnace image, image normalization, channel adjustment and image filling.

[0021] The second technical solution of the present invention is achieved by the following measures: a lightweight reduction furnace atomization detection device based on target detection technology, comprising:

[0022] An image acquisition and preprocessing module is used to acquire the image in the reduction furnace, preprocess the image in the reduction furnace, and obtain the preprocessed image in the reduction furnace;

[0023] A model building and training module is used to build and train a reduction furnace atomization detection model to obtain a trained reduction furnace atomization detection model;

[0024] The prediction module is used to input the preprocessed reduction furnace image into the trained reduction furnace atomization detection model to obtain the atomization detection result.

[0025] The third technical solution of the present invention is achieved through the following measures: an electronic device, including a memory and a processor, the memory stores a program that can be run on the processor, and when the processor executes the program, the above-mentioned lightweight reduction furnace atomization detection method based on target detection technology is implemented.

[0026] The fourth technical solution of the present invention is achieved through the following measures: a storage medium, which stores one or more programs, and the one or more programs can be executed by one or more processors to realize the above-mentioned lightweight reduction furnace atomization detection method based on target detection technology.

[0027] The present invention constructs and trains a reduction furnace atomization detection model to obtain a trained reduction furnace atomization detection model. The reduction furnace atomization detection model includes a ghost attention feature extraction network and a ghost attention aggregation feature pyramid network. The ghost attention feature extraction network and the ghost attention aggregation feature pyramid network replace the C2F module with the GS-C2F module on the basis of the feature extraction network and the path aggregation network of the YOLOv8 network, insert the CBAM module after the Backbone layer, and replace the CIOU function in the loss function with the WIOU loss function; the pre-processed reduction furnace operation image is input into the trained reduction furnace atomization detection model to obtain the atomization detection result, which solves the problem of large calculation amount and low accuracy. The present invention designs a G2F-GS module by using the idea of ​​reducing parameters and integrating the attention mechanism by using linear transformation. The G2F-GS module achieves higher parameter utilization efficiency and lower computational complexity by reducing the amount of parameters. The present invention applies the G2F-GS module to the feature extraction network of the YOLOv8 network, enriches the gradient flow of the feature extraction network, and reduces the amount of calculation. In addition, according to the G2F-GS module, a new feature pyramid network is designed based on the path aggregation network with feature pyramid network of YOLOv8 network, and the CBAM module is added, and the loss function is replaced with WIOU to enhance the multi-scale fusion of atomization features. While improving the accuracy of atomization detection, the present invention reduces the number of parameters and complexity of the model, improves the detection speed, and meets the real-time requirements of industrial detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a flow chart of a lightweight reduction furnace atomization detection method based on target detection technology according to an embodiment of the present invention.

[0029] Figure 2 It is a schematic diagram of a reduction furnace image of a lightweight reduction furnace atomization detection method based on target detection technology according to an embodiment of the present invention.

[0030] Figure 3 It is a schematic diagram of a reduction furnace atomization detection model of a lightweight reduction furnace atomization detection method based on target detection technology according to an embodiment of the present invention.

[0031] Figure 4 It is a structural schematic diagram of the GS-C2F module of the lightweight reduction furnace atomization detection method based on target detection technology in an embodiment of the present invention.

[0032] Figure 5 It is a structural schematic diagram of a lightweight reduction furnace atomization detection device based on target detection technology according to an embodiment of the present invention.

[0033] Figure 6 Schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0034] The present invention is not limited by the following embodiments, and specific implementation methods can be determined based on the technical solution of the present invention and actual conditions.

[0035] The present invention will be further described below in conjunction with embodiments:

[0036] Example 1: Figure 1 As shown, this embodiment provides a lightweight reduction furnace atomization detection method based on target detection technology, comprising the following steps:

[0037] Acquire the image inside the reduction furnace, preprocess the image inside the reduction furnace, and obtain the preprocessed image inside the reduction furnace;

[0038] A reduction furnace atomization detection model is constructed and trained to obtain a trained reduction furnace atomization detection model; in this embodiment, the constructed reduction furnace atomization detection model is a reduction furnace atomization detection model based on an improved YOLOv8 network; wherein the reduction furnace atomization detection model includes an input end, a Backbone, a Neck, and a Head, and Backbone replaces the C2F module with a GS-C2F module based on a feature extraction network of the YOLOv8 network and a path aggregation network with a feature pyramid network, adds a CBAM module after the Backbone layer, and replaces the CIOU function in the loss function with a WIOU loss function.

[0039] The preprocessed reduction furnace operation image is input into the trained reduction furnace atomization detection model to obtain the atomization detection result.

[0040] Preferably, the GS-C2F module includes a first CBS layer, a Spilt layer, a GhostBottleneck layer, a splicing layer, a second CBS layer and a SE layer connected in sequence. The input feature of the GS-C2F module passes through the first CBS layer to obtain the first feature. The Spilt layer divides the first feature into the second feature and the third feature in half according to the channel. The third feature passes through the GhostBottleneck layer to obtain the fourth feature. The second feature, the third feature and the fourth feature are dimensionally merged through the splicing layer, and finally calculated through the SE layer to obtain the output feature of the GS-C2F module.

[0041] As a preferred method, the GhostBottleneck layer integrates the Ghost convolution, divides the input into two parts: ordinary convolution and simple linear operation for feature extraction, and finally performs concat operation to fuse the features to obtain the final output; Ghost convolution first extracts the information features between channels through 1x 1 convolution, and then uses grouped convolution to generate a new feature map; in order to reduce the model calculation parameters, the traditional convolution is divided into two steps, firstly, ordinary convolution is used to generate a feature map with a smaller channel, and then a "cheap operation" is performed on this basis to perform a simple linear transformation; the formula of ordinary convolution is:

[0042] O=I*f

[0043] Among them, * represents the convolution operation, f is the convolution kernel, k*k is the convolution kernel size, O is the output feature map, and then L i,j Operate to produce Ghost feature map from each channel feature map of O;

[0044]

[0045] Among them, represents the i-th intrinsic feature map, represents the linear operation responsible for generating the j-th ghost map. The linear operation greatly reduces the large number of parameters generated by the model during the layer-by-layer convolution process. Finally, the feature maps obtained by ordinary convolution and linear transformation are spliced ​​together to get the final output.

[0046] Preferably, the first CBS layer and the second CBS layer both adopt a CBS structure, and the CBS structure includes a convolution layer with a convolution kernel size of 3×3, a batch normalization layer, and a silu activation function layer connected in sequence.

[0047] As a preference, the SE layer learns global context information for the input features of the GS-C2F module, thereby enhancing the model's focus on key features and effectively improving the model's expressiveness; the SE layer includes a Squeeze part and an Excitation part, wherein the Squeeze part uses a global average pooling operation to compress the spatial information of each feature map into a channel descriptor to reflect the overall activity level of the channel; wherein the Excitation part captures the dependencies between channels through two fully connected layers and learns the weights of each channel, thereby enhancing or suppressing specific feature responses. The mechanism is simple in design and easy to integrate into different CNN architectures.

[0048] Preferably, the preprocessing method includes adjusting the size of the reduction furnace image, image normalization, channel adjustment and image filling.

[0049] The lightweight reduction furnace atomization detection method based on target detection technology constructs a reduction furnace atomization detection model and trains it to obtain a trained reduction furnace atomization detection model. The reduction furnace atomization detection model includes a ghost attention feature extraction network and a ghost attention aggregation feature pyramid network. The ghost attention feature extraction network and the ghost attention aggregation feature pyramid network replace the C2F module with the GS-C2F module on the basis of the feature extraction network and the path aggregation network of the YOLOv8 network, insert the CBAM module after the Backbone layer, and replace the CIOU function in the loss function with the WIOU loss function; the preprocessed reduction furnace operation image is input into the trained reduction furnace atomization detection model to obtain the atomization detection result, which solves the problems of large calculation amount and low accuracy.

[0050] The lightweight reduction furnace atomization detection method based on target detection technology proposed in this embodiment designs the G2F-GS module by using linear transformation to reduce parameters and integrating the attention mechanism. The G2F-GS module achieves higher parameter utilization efficiency and lower computational complexity by reducing the amount of parameters. The G2F-GS module is applied to the feature extraction network of the YOLOv8 network, enriching the gradient flow of the feature extraction network and reducing the amount of calculation. In addition, based on the G2F-GS module, a new feature pyramid network is designed on the basis of the path aggregation network with a feature pyramid network of the YOLOv8 network, and the CBAM module is added, and the loss function is replaced by WIOU to enhance the multi-scale fusion of atomization features. While improving the accuracy of atomization detection, the parameter amount and complexity of the model are reduced, the detection speed is improved, and the real-time requirements of industrial detection are met.

[0051] Example 2: Figures 1 to 4 As shown, this embodiment provides a lightweight reduction furnace atomization detection method based on target detection technology, comprising the following steps:

[0052] S1. Acquire a reduction furnace image, and preprocess the reduction furnace image to obtain a preprocessed reduction furnace image. In this embodiment, the preprocessing method includes adjusting the size of the reduction furnace image, image normalization, channel adjustment, and image filling.

[0053] Specifically, in the training phase, an industrial camera is used to obtain reduction furnace images and annotate them, such as Figure 2As shown, the annotation format is txt format. In one embodiment, the reduction furnace image is preprocessed and cleaned, including removing useless atomization data, adjusting the size of the reduction furnace image taken by the industrial camera to 640×640 pixels and rotating it, and adjusting the contrast and other operations to obtain training data. In the inference stage, the size of the reduction furnace image taken by the industrial camera is adjusted to 640×640 pixels, and image normalization, channel adjustment, image filling and other operations are performed. In the present invention, the preprocessing process can be adjusted according to actual needs.

[0054] S2. Build and train a reduction furnace atomization detection model based on the improved YOLOv8 network to obtain a trained reduction furnace atomization detection model. The reduction furnace atomization detection model includes a ghost attention feature extraction network and a ghost attention aggregation feature pyramid network. The ghost attention feature extraction network and the ghost attention aggregation feature pyramid network replace the C2F module with the GS-C2F module on the basis of the feature extraction network and the path aggregation network of the YOLOv8 network, add the CBAM module after the Backbone layer, and replace the CIOU function in the loss function with the WIOU loss function.

[0055] In this embodiment, the GS-C2F module includes a first CBS layer, a Spilt layer, a GhostBottleneck layer, a splicing layer, a second CBS layer, and a SE layer connected in sequence. The input feature of the GS-C2F module passes through the first CBS layer to obtain the first feature. The Spilt layer divides the first feature into the second feature and the third feature in half according to the channel. The third feature passes through the GhostBottleneck layer to obtain the fourth feature. The second feature, the third feature, and the fourth feature are dimensionally merged through the splicing layer, and finally calculated through the SE layer to obtain the output feature of the GS-C2F module.

[0056] Specifically, refer to Figure 3 , the reduction furnace atomization detection model based on the improved YOLOv8 network is improved on the basis of the YOLOv8 network. In one embodiment, the reduction furnace atomization detection model is improved on the basis of YOLOv8l. In other embodiments, other types of YOLOv8 networks can also be used. The GS-C2F module replaces the C2F module to design a lightweight ghost attention feature extraction network (GSC2F-net) and an efficient ghost attention aggregation feature pyramid network (GSC2F-FPN). At the same time, the CBAM module is inserted to replace the CIOU function in the loss function with the WIOU loss function. Therefore, the reduction furnace atomization detection model belongs to a lightweight target detection model.

[0057] Furthermore, the ghost attention feature extraction network (GSC2F-net) extracts feature information from the reduction furnace image through the GS-C2F module and outputs feature maps M1, M2, and M3.

[0058] Specifically, the ghost attention feature extraction network (GSC2F-net) includes: the first layer is a CBS module, which uses a convolution layer with 64 channels and a convolution kernel size of 3×3 to perform a convolution operation on the input reduction furnace image, obtains a feature map of 64 channels, and reduces it to half of the original size. The second layer is a CBS module, which uses a convolution layer with 128 channels and a convolution kernel size of 3×3 to perform a further convolution operation on the feature map, obtains a feature map of 128 channels, and reduces it to half of the original size again. The CBS modules all adopt a CBS structure, that is, a batch normalization layer (BN) and a silu activation function layer are connected in sequence after the convolution layer. The third layer is a C2F module, which fuses feature maps of different scales, and the number of channels is 128. The fourth layer is a downsampling module, which uses 256 channels to perform a convolution operation on the feature map, obtains a feature map of 256 channels, and reduces it to half of the original size again. The fifth layer is a C2F module, which further fuses the feature map, with 256 channels and unchanged size, and outputs the feature map M3. The sixth layer is a downsampling module, which uses 512 channels to perform convolution operations on the feature map M3, obtains a feature map of 512 channels, and reduces it to half its original size again. The seventh layer is a GS-C2F module, which further fuses the feature map, with 512 channels, and outputs the feature map M2. The eighth layer is a downsampling module, which uses 1024 channels to perform convolution operations on the feature map M2, obtains a feature map of 512 channels, and reduces it to half its original size again. The ninth layer is a GS-C2F module, which fuses feature maps of different scales, with 1024 channels. The tenth layer is the SPPF module. The SPPF module uses a 1024-channel spatial pyramid pooling operation to extract features from different scales, capture the detailed information of the target object at different scales, and output the feature map M1.

[0059] like Figure 4As shown in the figure, the GS-C2F module enriches the gradient flow by integrating the changes of gradient information into the feature map from beginning to end, which can ensure accuracy while reducing the amount of calculation, making it suitable for running on resource-constrained devices. The specific structure is as follows: the first CBS layer, the Spilt layer, the GhostBottleneck layer, the splicing layer, the second CBS layer, and the SE layer. The input feature of the GS-C2F module passes through the first CBS layer to obtain the first feature. The Spilt layer divides the first feature into the second feature and the third feature in half according to the channel. The third feature passes through the GhostBottleneck layer to obtain the fourth feature. The second feature, the third feature, and the fourth feature are merged by dimension through the splicing layer. Finally, the SE layer is used for calculation to obtain the output feature of the GS-C2F module. The first feature is obtained by the input feature of the GS-C2F module through the first CBS layer. After that, the first feature is divided into the second feature and the third feature in half according to the number of channels after the dimension splitting layer. The third feature is passed through the GhostBottleneck layer to obtain the fourth feature. The splicing layer merges the feature maps containing different information, namely the second feature, the third feature, and the fourth feature dimensions, and passes through the fourth CBS layer. Finally, it is calculated through the SE layer to obtain the output feature of the GS-C2F module.

[0060] In this embodiment, GhostBottleneck uses Ghost convolution to divide the input into two parts: ordinary convolution and simple linear operation for feature extraction, and finally performs concat operation to fuse the features to obtain the final output. Ghost convolution first extracts the information features between channels through 1x 1 convolution, and then uses grouped convolution to generate a new feature map. In order to reduce the model calculation parameters, the traditional convolution is divided into two steps. First, ordinary convolution is used to generate a feature map with a smaller channel, and then a "cheap operation" is performed on this basis to perform a simple linear transformation. The first step of ordinary convolution can be formulated as:

[0061] O=I*f

[0062] Among them, * represents the convolution operation, f is the convolution kernel, k*k is the convolution kernel size, O is the output feature map, and then L i,j Operation is used to produce a Ghost feature map from each channel feature map of O.

[0063]

[0064] where represents the i-th intrinsic feature map, and represents the linear operation responsible for generating the j-th ghost map. The linear operation greatly reduces the large number of parameters generated by the model during the layer-by-layer convolution process. Finally, the feature maps obtained by ordinary convolution and linear transformation are concatenated together to obtain the final output.

[0065] In this embodiment, GhostBottleneck also uses the SE attention mechanism, which is a mechanism that effectively improves the model's expressiveness by learning global context information to enhance the model's attention to key features. It is divided into two parts. The Squeeze part uses a global average pooling operation to compress the spatial information of each feature map into a channel descriptor to reflect the overall activity level of the channel. The Excitation part captures the dependencies between channels through two fully connected layers and learns the weights of each channel to enhance or suppress specific feature responses.

[0066] In this embodiment, the CBS structure includes a convolution layer with a convolution kernel size of 3×3, a batch normalization layer, and a silu activation function layer connected in sequence. The specific fusion process of the ghost attention aggregate feature pyramid network (GSC2F-FPN) is as follows:

[0067] (a) M1 feature map and M2 feature map after upsampling module Figure 1 The input concatenation layer performs dimension concatenation operation to obtain feature map T2;

[0068] (b) Feature map T2 after passing through the C2F module and the CBAM module and after the upsampling module and M3 feature map Figure 1 The input concatenation layer performs dimension concatenation operation to obtain feature map T3;

[0069] (c) Feature map T2 and T3 after the C2F module and downsampling module are input into the concatenation layer for dimensional concatenation to obtain feature map T4;

[0070] (d) The feature map M1 and T4 after the C2F module and the downsampling module are input into the concatenation layer for dimensional concatenation to obtain the feature map T1;

[0071] (e) Feature map R3 is obtained by passing feature map T3 through the C2F module; feature map R2 is obtained by passing feature map T4 through the C2F module; feature map R1 is obtained by passing feature map T1 through the GS-C2F module.

[0072] The specific mathematical expression is:

[0073] R 3 =C2F(T 3 )

[0074] R 2 =C2F(T 4 )

[0075] R 1 =GS-C2F(T 1

[0076] T 3 =Concat{M3 ,upasmpling[C2F(T 2 )]}

[0077] T 2 =Concat{M 2 ,upasmpling(M 1 )}

[0078] T 4 =Concat{downasmpling[C2F(T 3 )],C2F(T 2 )}

[0079] T 1 =Concat{downasmpling[C2F(T 4 )],M 1}

[0080] Where Concat represents the dimensional concatenation operation corresponding to the concatenation layer, upsampling represents the upsampling operation corresponding to the upsampling module, downsampling represents the downsampling operation corresponding to the downsampling module, C2F represents the function corresponding to the C2F module, and GS-C2F represents the function corresponding to the GS-C2F module.

[0081] In this embodiment, in terms of category classification loss, VFL Loss is used as the classification loss. VFL Loss measures the difference between the probability distribution predicted by the model and the actual label. Its formula is as follows:

[0082]

[0083] Among them, b is the IoU between bbox and gt, a is the probability, when b is a positive sample, the classification loss is calculated as BCE (binary cross entropy) loss, when it is a negative sample, it is FL.

[0084] In this embodiment, in terms of border regression loss, DFL Loss + WIOU Loss is used as the border regression loss. Specifically, after YOLOv8 introduces the idea of ​​Center-based methods (based on the center point) of Anchor-Free, the model changes from outputting offest to ltrb = left, top, right, bottom, that is, the distance from the four borders of the predicted box to the center point. In order to cooperate with Anchor-Free and improve generalization, DFL loss is added in YOLOv8, and its formula is as follows:

[0085] DFL(S i ,S i+1 )=-((yi+1 -y)log(S i )+(yy i )log(S i+1 ))

[0086] After calculating ltrb, DFL uses cross entropy to allow the network to focus on the target location and the distribution of neighboring areas more quickly.

[0087] Since the training data necessarily includes low-quality examples, and geometric factors such as distance and aspect ratio will aggravate the impact of such low-quality examples, thereby reducing the generalization performance of the model. A good loss function should reduce the penalty of geometric factors when the anchor box and the true box basically overlap, and less intervention during training can enable the model to have better generalization ability. Therefore, the distance attention is constructed based on the distance metric, and WIoUv1 with a two-layer attention mechanism is obtained, and its formula is as follows:

[0088] L WIoUv1 =R WIoU L IoU

[0089]

[0090] Where Wg, Hg are the sizes of the smallest enclosing box. WIoU The gradients that hinder convergence are generated, and Wg, Hg are separated from the computational graph (the superscript * indicates this operation). Because it effectively eliminates the factors that hinder convergence, no new metrics such as aspect ratio are introduced. Therefore, WIoU solves the problem of CIoU when the aspect ratio is the same.

[0091] Wise-IoUv3 introduces a dynamic non-monotonic attention mechanism to replace the static monotonic attention mechanism. This mechanism uses "outliers" instead of IoU to evaluate the quality of anchor boxes, and implements a wise gradient gain allocation strategy. This strategy can reduce the competitiveness of high-quality anchor boxes and reduce the harmful gradients generated by low-quality samples, so that WIoU can give priority to anchor boxes of ordinary quality and improve detector performance. The formula is as follows:

[0092]

[0093]

[0094] Using β to construct a non-monotonic focusing coefficient and apply it to WIoU v1, we obtain WIoU v3 with dynamic non-monotonic FM, which can dynamically optimize the loss to weight small objects to improve the detection performance of the model.

[0095] The reduction furnace atomization detection model is trained using the Adaw optimizer with a learning rate of 0.001. The training data is divided into training set, validation set and test set according to the ratio of 8:1:1. The reduction furnace images in the training set are input into the reduction furnace atomization detection model, and the trained reduction furnace atomization detection model is deployed and applied to realize the detection of reduction furnace atomization.

[0096] S3. Input the preprocessed reduction furnace image into the trained atomization detection model to obtain the atomization detection result.

[0097] Specifically, in the inference stage, the preprocessed reduction furnace image is passed through a trained atomization detection model to predict the atomization detection result, which includes the atomization category and the atomization position.

[0098] The lightweight reduction furnace atomization detection method based on target detection technology proposed in this embodiment designs the G2F-GS module by using linear transformation to reduce parameters and integrating the attention mechanism. The G2F-GS module achieves higher parameter utilization efficiency and lower computational complexity by reducing the amount of parameters. The G2F-GS module is applied to the feature extraction network of the YOLOv8 network, enriching the gradient flow of the feature extraction network and reducing the amount of calculation. In addition, based on the G2F-GS module, a new feature pyramid network is designed on the basis of the path aggregation network with a feature pyramid network of the YOLOv8 network, and the CBAM module is added, and the loss function is replaced by WIOU to enhance the multi-scale fusion of atomization features. While improving the accuracy of atomization detection, the parameter amount and complexity of the model are reduced, the detection speed is improved, and the real-time requirements of industrial detection are met.

[0099] Example 3: Figure 5 As shown, this embodiment provides a lightweight reduction furnace atomization detection device based on target detection technology, which can be specifically applied to various electronic devices, including:

[0100] An image acquisition and preprocessing module is used to acquire a reduction furnace image, preprocess the reduction furnace image, and obtain a preprocessed reduction furnace image;

[0101] A model construction training module is used to construct and train a reduction furnace atomization detection model based on an improved YOLOv8 network to obtain a trained reduction furnace atomization detection model; wherein the constructed reduction furnace atomization detection model is a reduction furnace atomization detection model based on an improved YOLOv8 network; wherein the reduction furnace atomization detection model includes a ghost attention feature extraction network and a ghost attention aggregation feature pyramid network, and the ghost attention feature extraction network and the ghost attention aggregation feature pyramid network replace the C2F module with the GS-C2F module on the basis of the feature extraction network and the path aggregation network of the YOLOv8 network, insert the CBAM module after the Backbone layer, and replace the CIOU function in the loss function with the WIOU loss function.

[0102] The prediction module is used to input the preprocessed reduction furnace image into the trained reduction furnace atomization detection model to obtain the atomization detection result.

[0103] The lightweight reduction furnace atomization detection device based on target detection technology proposed in this embodiment designs a G2F-GS module by using linear transformation to reduce parameters and integrating the attention mechanism. The G2F-GS module achieves higher parameter utilization efficiency and lower computational complexity by reducing the amount of parameters. The G2F-GS module is applied to the feature extraction network of the YOLOv8 network, enriching the gradient flow of the feature extraction network and reducing the amount of calculation. In addition, based on the G2F-GS module, a new feature pyramid network is designed on the basis of the path aggregation network with a feature pyramid network of the YOLOv8 network, and a CBAM module is added, and the loss function is replaced with WIOU to enhance the multi-scale fusion of atomization features. While improving the accuracy of atomization detection, the parameter amount and complexity of the model are reduced, the detection speed is improved, and the real-time requirements of industrial detection are met.

[0104] Example 4: Figure 6 As shown, this embodiment provides an electronic device, which includes a memory, a processor, a communication interface and a bus. The memory stores a program that can be run on the processor. When the processor executes the program, the lightweight reduction furnace atomization detection method based on target detection technology in the above embodiment is implemented.

[0105] like Figure 6As shown, the electronic device of this embodiment may include a processor 1001 and a memory 1002; wherein the memory 1002 is used to store computer-executable instructions; the processor 1001 is used to execute the computer-executable instructions stored in the memory to implement the various steps of the lightweight reduction furnace atomization detection method based on target detection technology in the above embodiment, and the details can be referred to the relevant description in the above method embodiment. Optionally, the memory 1002 can be either independent or integrated with the processor 1001. When the memory 1002 is independently provided, the electronic device also includes a bus 1003 for connecting the memory 1002 and the processor 1001.

[0106] The processor may be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, field programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or a combination of the above chips.

[0107] The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs, non-transient computer executable programs and units, such as the corresponding program units in the above method embodiments of the present invention. The processor executes various functional applications and data processing of the processor by running the non-transient software programs, instructions and modules stored in the memory, thereby realizing the lightweight reduction furnace atomization detection method based on target detection technology in the above embodiment.

[0108] The memory may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required by at least one function; the data storage area may store data created by the processor, etc. In addition, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. The memory may optionally include a memory remotely disposed relative to the processor, and these remote memories may be connected to the processor via a network. One or more programs are stored in the memory, and when executed by the processor, the lightweight reduction furnace atomization detection method based on the target detection technology in the above-mentioned embodiment is executed.

[0109] Embodiment 5: This embodiment provides a storage medium for computer-readable storage, the storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the lightweight reduction furnace atomization detection method based on target detection technology as in the above embodiment.

[0110] The storage medium may be an internal storage unit of the electronic device, such as a hard disk or memory of the electronic device. The storage medium may also be an external storage device of the electronic device, such as a plug-in hard disk, smart memory card, secure digital card, flash memory card, etc. equipped on the electronic device.

[0111] The above technical features constitute the embodiments of the present invention, which have strong adaptability and implementation effect. Non-essential technical features can be added or reduced according to actual needs to meet the requirements of different situations.

Claims

1. A lightweight reduction furnace atomization detection method based on target detection technology, characterized in that The following steps are involved: Acquire the image inside the reduction furnace, preprocess the image inside the reduction furnace, and obtain the preprocessed image inside the reduction furnace; Constructing and training a reduction furnace atomization detection model to obtain a trained reduction furnace atomization detection model; The preprocessed reduction furnace operation image is input into the trained reduction furnace atomization detection model to obtain the atomization detection result.

2. The method for detecting atomization of a lightweight reduction furnace based on target detection technology according to claim 1 is characterized in that The reduction furnace atomization detection model includes input end, Backbone, Neck, and Head. Backbone replaces the C2F module with the GS-C2F module based on the feature extraction network of the YOLOv8 network and the path aggregation network with the feature pyramid network. The CBAM module is added after the Backbone layer, and the CIOU function in the loss function is replaced by the WIOU loss function.

3. The method for detecting atomization of a lightweight reduction furnace based on target detection technology according to claim 2 is characterized in that The GS-C2F module includes the first CBS layer, Spilt layer, GhostBottleneck layer, splicing layer, second CBS layer and SE layer connected in sequence. The input feature of the GS-C2F module passes through the first CBS layer to obtain the first feature. The Spilt layer divides the first feature into the second feature and the third feature in half according to the channel. The third feature passes through the GhostBottleneck layer to obtain the fourth feature. The second feature, the third feature and the fourth feature are dimensionally merged through the splicing layer. Finally, they are calculated through the SE layer to obtain the output feature of the GS-C2F module.

4. The method for detecting atomization of a lightweight reduction furnace based on target detection technology according to claim 3 is characterized in that The GhostBottleneck layer integrates the Ghost convolution, divides the input into two parts: ordinary convolution and simple linear operation for feature extraction, and finally performs concat operation to fuse the features to obtain the final output; Ghost convolution first extracts the information features between channels through convolution, and then uses grouped convolution to generate a new feature map; ordinary convolution is used to generate a feature map with a smaller channel, and then a linear transformation is performed; the formula for ordinary convolution is: O=I*f Among them, * represents the convolution operation, f is the convolution kernel, k*k is the convolution kernel size, O is the output feature map, and then L i,j Operate to produce Ghost feature map from each channel feature map of O; Among them, represents the i-th intrinsic feature map, represents the linear operation responsible for generating the j-th ghost map; finally, the feature maps obtained by ordinary convolution and linear transformation are spliced ​​together to get the final output.

5. The lightweight reduction furnace atomization detection method based on target detection technology according to claim 3 or 4, characterized in that The first CBS layer and the second CBS layer both adopt the CBS structure, which includes a convolutional layer, a batch normalization layer, and a silu activation function layer connected in sequence.

6. The method for detecting atomization of a lightweight reduction furnace based on target detection technology according to claim 3 or 4, characterized in that The SE layer learns the global context information of the input features of the GS-C2F module; the SE layer includes a Squeeze part and an Excitation part. The Squeeze part uses a global average pooling operation to compress the spatial information of each feature map into a channel descriptor to reflect the overall activity level of the channel; the Excitation part captures the dependencies between channels through two fully connected layers and learns the weights of each channel to enhance or suppress specific feature responses.

7. The method for detecting atomization of a lightweight reduction furnace based on target detection technology according to claim 1, 2, 3 or 4, characterized in that The preprocessing methods include adjusting the size of the reduction furnace image, image normalization, channel adjustment and image filling.

8. A lightweight reduction furnace atomization detection device based on target detection technology, characterized in that include: An image acquisition and preprocessing module is used to acquire the image in the reduction furnace, preprocess the image in the reduction furnace, and obtain the preprocessed image in the reduction furnace; A model building and training module is used to build and train a reduction furnace atomization detection model to obtain a trained reduction furnace atomization detection model; The prediction module is used to input the preprocessed reduction furnace image into the trained reduction furnace atomization detection model to obtain the atomization detection result.

9. An electronic device comprising a memory and a processor, wherein the memory stores a program that can be run on the processor, characterized in that: When the processor executes the program, the lightweight reduction furnace atomization detection method based on target detection technology as described in any one of claims 1 to 7 is implemented.

10. A storage medium, characterized in that: The storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the lightweight reduction furnace atomization detection method based on target detection technology as described in any one of claims 1 to 7.

Images