Image data acquisition method under dangerous condition

Abstract

The invention discloses an image data acquisition method under a dangerous condition, and relates to the technical field of image recognition. The method comprises the steps of obtaining an original image of a target area, and performing preprocessing operation on the original image to obtain an initial image; substituting the initial image into a dangerous scene recognition model for recognition to obtain a recognition feature value; judging scene danger information of the monitoring image according to the identification characteristic value; continuously capturing the monitoring image of the target area through the scene danger information to obtain a danger image; and transmitting the dangerous image to a user terminal to complete output. According to the method, the initial image is identified through the dangerous scene identification model, scene dangerous information can be judged based on the identification characteristic value, and the dangerous image is continuously captured. According to the accurate recognition and capture mechanism, the non-dangerous redundant images can be effectively filtered, it is ensured that the dangerous images which are finally conveyed to the user terminal are really valuable dangerous images, the accuracy of capturing the dangerous images is improved, and then the overall efficiency of obtaining the images is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image recognition, and particularly relates to an image data acquisition method under a dangerous condition. BACKGROUND

[0002] With the development of image recognition technology, image data acquisition under a dangerous condition (such as a fire scene, a chemical explosion area, a geological disaster scene, etc.) has always been a difficult problem to overcome. Traditional image acquisition equipment is often unable to work stably under a dangerous condition due to various factors, resulting in that the on-site image data cannot be acquired in time and accurately, which brings great obstacles to subsequent disaster assessment, rescue decision-making and the like.

[0003] The traditional image data acquisition method under a dangerous condition mainly relies on manual visual detection. On the one hand, the detection efficiency is extremely low: manual observation is required after arriving at the scene, and it is impossible to quickly acquire images of a large range of dangerous scenes, which seriously lags behind the time requirement of disaster disposal, for example, the fire spreads fast in a fire scene, and manual visual detection is difficult to complete the acquisition of image data of the whole area in a short time. On the other hand, the data precision is seriously insufficient: manual visual detection can only rely on subjective visual judgment, and cannot quantify the key information in the image, so that the precision of the data cannot be guaranteed, and the overall efficiency of acquiring image data is low. SUMMARY

[0004] The purpose of the present application is to solve the problem of low overall efficiency of acquiring image data, and to propose an image data acquisition method under a dangerous condition.

[0005] In the first aspect of the present application, an image data acquisition method under a dangerous condition is first proposed, and the method comprises: acquiring an original image of a target area, and performing a preprocessing operation on the original image to obtain an initial image; putting the initial image into a dangerous scene recognition model to obtain a recognition feature value; judging the scene danger information of the monitoring image according to the recognition feature value, and continuously capturing a dangerous image of the target area through the scene danger information; delivering the dangerous image to a user terminal to complete output.

[0006] Optionally, the preprocessing operation on the original image to obtain the initial image comprises: performing a 3x3 convolution operation on the original image to obtain a first convolution tensor; performing average pooling and maximum pooling on the first convolution tensor respectively to obtain a first average tensor and a first maximum tensor; splicing the first average tensor and the first maximum tensor to obtain a spliced tensor; Substituting the spliced ​​vector into a multilayer perceptron yields a channel feature map; The first feature map is obtained by multiplying the monitoring image and the channel feature map; After performing a 3×3 convolution operation on the first feature map, a sigmoid activation function is input to obtain a spatial feature map. The initial image is obtained by multiplying the channel feature map, spatial feature map, and first feature map.

[0007] Optionally, the principle process of the hazardous scene recognition model includes: The initial image is processed sequentially through a first convolutional layer, a second convolutional layer, a first star layer, a third convolutional layer, and a second star layer to obtain a first feature map; The first feature map is processed sequentially through the fourth convolutional layer and the third star layer to obtain the second feature map; The second feature map is processed sequentially through the fifth convolutional layer, the fourth star layer, the feature enhancement layer, and the SPPF layer to obtain the third feature map; The third feature map is upsampled and then fused with the second feature map to obtain the fourth feature map. The fourth feature map is processed through the C2f_ML layer to obtain the fifth feature map; the fifth feature map is then upsampled and fused with the first feature map to obtain the sixth feature map; The sixth feature map is processed through a C2f_ML layer and a sixth convolutional layer to obtain the seventh feature map; the seventh feature map and the fifth feature map are then fused to obtain the eighth feature map; The eighth feature map is processed through a C2f_ML layer and a seventh convolutional layer to obtain the ninth feature map; the ninth feature map and the third feature map are then fused to obtain the ninth feature map. The ninth feature map is processed through the C2f_ML layer to obtain the target feature map; The target feature map is used to search a preset database to determine the identification feature value.

[0008] Optionally, the working principle of the star-shaped layer includes: Use the feature map of the input star layer as the input feature; The input features are processed through the DWConv layer to obtain the first feature; Substituting the first feature into the first fully connected layer and the second fully connected layer in sequence, we obtain the first fully connected feature and the second fully connected feature, respectively. The second feature is obtained by multiplying the first fully connected feature and the second fully connected layer feature; The second feature is substituted into the third fully connected layer to obtain the third feature; The third feature is substituted into the DWConv layer for processing to obtain the target feature.

[0009] Optionally, the working principle of the C2f_ML layer includes: Use the feature map of the input C2f_ML layer as the input feature tensor; The first feature tensor is obtained by substituting the input feature tensor into the Split layer; The first feature tensor is substituted into the first Bottleneck layer to obtain the second feature tensor; the second feature tensor is substituted into the second Bottleneck layer to obtain the third feature tensor. Substitute the first feature tensor, the second feature tensor, and the third feature tensor into the Concat layer to obtain the fused feature tensor; The fused feature tensor is substituted into the convolutional layer to obtain the output feature tensor.

[0010] Optionally, the working principle of the Bottleneck layer includes: The Bottleneck layer is composed of two convolutional layers and an ML layer connected in sequence. The specific working principle of the ML layer includes: The feature map input to the ML layer is processed by a local attention pooling layer to obtain attention pooling features; The attention pooling features are obtained by performing a 1×1 convolution operation on the attention pooling features; The attention pooling features and the attention pooling features are fused to obtain the fused features; The local target features are obtained by multiplying the feature map and the fused features.

[0011] Optionally, the principle process of the feature enhancement layer includes: Where Y1 is the input of the feature enhancement layer, Y2, Y3, Y4 and Y5 are the outputs of the four two-dimensional convolution branches respectively; Z1 is the attention weight; Z is the output of the feature enhancement module; f is the operator, the symbol ⊗ indicates element-wise multiplication, the superscript indicates the kernel size, and the subscript Conv2d indicates two-dimensional convolution.

[0012] Optionally, determining the scene hazard information of the monitored image based on the identified feature values ​​includes: If the identified feature value is greater than the preset feature threshold, the scene of the detected image is in a dangerous state.

[0013] Optionally, the step of determining the scene hazard information of the monitored image based on the identification feature value further includes: If the identified feature value is less than or equal to the preset feature threshold, the scene of the detected image is in a safe state.

[0014] The beneficial effects of this invention are: This invention proposes a method for acquiring image data under hazardous conditions. The method involves acquiring the original image of the target area, preprocessing the original image to obtain an initial image, inputting the initial image into a hazardous scene recognition model to obtain recognition feature values, determining the scene hazard information of the monitored image based on the recognition feature values, continuously capturing hazardous images of the target area based on the scene hazard information, and finally transmitting the hazardous images to the user terminal for output. This invention uses a hazardous scene recognition model to identify the initial image, accurately determining scene hazard information based on recognition feature values, and then selectively and continuously capturing hazardous images. This precise identification and capture mechanism effectively filters out redundant images without hazard, ensuring that only truly valuable hazardous images are ultimately delivered to the user terminal, improving the accuracy of hazardous image capture and thus increasing the overall efficiency of image acquisition. Attached Figure Description

[0015] The invention will now be further described with reference to the accompanying drawings.

[0016] Figure 1 A flowchart illustrating a method for acquiring image data under hazardous conditions, as provided in an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The term "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and B can represent: A alone, A and B simultaneously, and B alone. Furthermore, descriptions involving "first," "second," etc., in this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" or "second" can explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0018] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] This invention provides a method for acquiring image data under dangerous conditions. See also...Figure 1 , Figure 1 A flowchart illustrating a method for acquiring image data under hazardous conditions, as provided in an embodiment of the present invention. The method includes the following steps: Obtain the original image of the target region, and perform preprocessing operations on the original image to obtain the initial image; The initial image is substituted into the hazardous scene recognition model to obtain recognition feature values; Based on the identified feature values, the scene hazard information of the monitoring image is determined; through the scene hazard information, the monitoring image of the target area is continuously captured to obtain dangerous images; The dangerous image is sent to the user terminal to complete the output.

[0020] The image data acquisition method under dangerous conditions provided in this invention identifies the initial image using a dangerous scene recognition model. Based on the identified feature values, it accurately determines the danger information of the scene and then continuously captures dangerous images. This precise identification and capture mechanism effectively filters out redundant images without danger, ensuring that only truly valuable dangerous images are ultimately delivered to the user terminal. This improves the accuracy of capturing dangerous images and thus enhances the overall efficiency of image acquisition.

[0021] In one implementation, preprocessing the original image to obtain the initial image includes: Perform a 3×3 convolution operation on the original image to obtain the first convolution tensor; The first convolutional tensor is subjected to average pooling and max pooling respectively to obtain the first average tensor and the first max tensor; The first average tensor and the first maximum tensor are concatenated to obtain the concatenated tensor; Substituting the concatenated vectors into a multilayer perceptron yields the channel feature map. The first feature map is obtained by multiplying the monitoring image and the channel feature map; After performing a 3×3 convolution operation on the first feature map, the sigmoid activation function is passed in to obtain the spatial feature map. The initial image is obtained by multiplying the channel feature map, spatial feature map, and first feature map.

[0022] In one implementation, to improve the efficiency of dangerous scene recognition, the image preprocessing process achieves a balance between comprehensive feature extraction, accurate recognition, and efficient processing through a multi-stage design. The process first uses 3×3 convolution operations combined with average pooling and max pooling to extract multi-dimensional low-level features such as edges and textures from the original image. Average pooling preserves overall trend information, while max pooling highlights salient features. The two are then concatenated and processed by a multilayer perceptron to mine deep features in the channel dimension, providing rich input for subsequent recognition models. Next, the monitored image is multiplied with the channel feature map and combined with the spatial feature map to complete multi-dimensional feature fusion, accurately capturing the spatial distribution of dangerous scenes and the correlation between channel features. For example, in fire scenes, it can not only identify spatial visual features such as flame color and shape, but also enhance the salience of the flame area, helping the model to more accurately distinguish different dangerous scenes such as fires, explosions, and geological disasters, reducing misjudgments. At the same time, the process uses mature image processing modules such as convolution, pooling, and multilayer perceptrons to control computational complexity while ensuring the feature extraction effect, ensuring efficient operation on image acquisition equipment under dangerous conditions, and avoiding the impact of excessive preprocessing time on the timeliness of subsequent dangerous image capture and output.

[0023] In one implementation method, the principle and process of the hazardous scene recognition model include: The initial image is processed sequentially through the first convolutional layer, the second convolutional layer, the first star layer, the third convolutional layer, and the second star layer to obtain the first feature map; The first feature map is processed sequentially through the fourth convolutional layer and the third star layer to obtain the second feature map; The second feature map is processed sequentially through the fifth convolutional layer, the fourth star layer, the feature enhancement layer, and the SPPF layer to obtain the third feature map; The third feature map is upsampled and then fused with the second feature map to obtain the fourth feature map. The fourth feature map is processed through the C2f_ML layer to obtain the fifth feature map; the fifth feature map is then upsampled and fused with the first feature map to obtain the sixth feature map. The sixth feature map is processed through the C2f_ML layer and the sixth convolutional layer to obtain the seventh feature map; the seventh feature map and the fifth feature map are fused to obtain the eighth feature map; The eighth feature map is processed through the C2f_ML layer and the seventh convolutional layer to obtain the ninth feature map; the ninth feature map and the third feature map are then fused to obtain the ninth feature map. The ninth feature map is processed through the C2f_ML layer to obtain the target feature map; The target feature map is used to search a preset database to determine the identification feature value.

[0024] In one implementation, the hazardous scene recognition model significantly improves recognition performance through multi-stage feature processing. The model processes the initial image using convolutional layers from the first to the seventh layer, combined with four sets of star-shaped layers. Features are extracted layer by layer from the bottom to the top, with the star-shaped layers enhancing key feature capture and avoiding biases caused by insufficient early feature extraction, laying the foundation for subsequent optimization. The model performs cross-stage feature fusion multiple times: the upsampled third and fifth feature maps are fused with the second and first feature maps, respectively, and the seventh and ninth feature maps are then fused with the fifth and third feature maps. This process integrates features from both shallow and deep layers, preserving details such as local textures and edges in hazardous scenes while incorporating semantic information such as overall shape and category association, allowing the model to perceive complex scenes such as fires and explosions more comprehensively. Simultaneously, the feature enhancement layer, SPPF layer, and C2f_ML layer work synergistically: the former enhances key feature signals, the SPPF layer reduces computational load and efficiently enhances global features, and the latter optimizes feature transfer efficiency, balancing processing quality and running speed to avoid recognition delays. Ultimately, upsampling combined with fusion achieves feature complementarity. The generated target feature map is combined with the recognition value from the preset database, which greatly improves the recognition accuracy, reduces missed and false judgments, and provides support for obtaining image data.

[0025] In one implementation, the working principle of the star topology includes: Use the feature map of the input star layer as the input feature; The input features are processed through the DWConv layer to obtain the first feature; Substituting the first feature into the first fully connected layer and the second fully connected layer in sequence, we obtain the first fully connected feature and the second fully connected feature, respectively. The second feature is obtained by multiplying the first fully connected feature and the second fully connected layer feature; The second feature is substituted into the third fully connected layer to obtain the third feature; The third feature is substituted into the DWConv layer for processing to obtain the target feature.

[0026] In one implementation, the working principle of the C2f_ML layer includes: Use the feature map of the input C2f_ML layer as the input feature tensor; The first feature tensor is obtained by substituting the input feature tensor into the Split layer; Substitute the first feature tensor into the first Bottleneck layer to obtain the second feature tensor; substitute the second feature tensor into the second Bottleneck layer to obtain the third feature tensor; Substitute the first feature tensor, the second feature tensor, and the third feature tensor into the Concat layer to obtain the fused feature tensor; The fused feature tensor is substituted into the convolutional layer to obtain the output feature tensor.

[0027] In one implementation, the working principle of the Bottleneck layer includes: The Bottleneck layer is composed of two convolutional layers and an ML layer connected in sequence. The specific working principle of the ML layer includes: The feature map input to the ML layer is processed by a local attention pooling layer to obtain attention pooling features; Attention pooling features are obtained by performing a 1×1 convolution operation on the attention pooling features; The attention pooling features are fused together to obtain the fused features; The local target features are obtained by multiplying the feature map and the fused features.

[0028] In one implementation, the principle and process of the feature enhancement layer include: Where Y1 is the input of the feature enhancement layer, Y2, Y3, Y4 and Y5 are the outputs of the four two-dimensional convolution branches respectively; Z1 is the attention weight; Z is the output of the feature enhancement module; f is the operator, the symbol ⊗ indicates element-wise multiplication, the superscript indicates the kernel size, and the subscript Conv2d indicates two-dimensional convolution.

[0029] In one implementation, determining the scene hazard information of the monitored image based on the identified feature values ​​includes: If the identified feature value is greater than the preset feature threshold, the scene of the detected image is in a dangerous state.

[0030] In one implementation, determining the scene hazard information of the monitored image based on the identified feature values ​​further includes: If the identified feature value is less than or equal to the preset feature threshold, the scene of the detected image is in a safe state.

[0031] Specifically, the preset feature thresholds are obtained based on the staff's historical experience; scenario hazard information includes explosions, leaks (toxic gases, chemicals), collapses, and rail transit malfunctions in industrial production, as well as falls from heights and mechanical injuries in construction; equipment malfunctions: such as electrical equipment short circuits, special equipment (elevators, boilers) malfunctions, fire-fighting facilities (fire extinguishers, sprinkler systems) failures, and traffic light malfunctions; facility defects: such as cracked building walls, damaged bridges and roads, loose guardrails, missing manhole covers, and aging power lines.

[0032] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for acquiring image data under dangerous conditions, characterized in that, The method includes: Obtain the original image of the target region, and perform preprocessing operations on the original image to obtain an initial image; The initial image is substituted into the hazardous scene recognition model to obtain recognition feature values; Based on the identified feature values, the scene hazard information of the monitoring image is determined; through the scene hazard information, the monitoring image of the target area is continuously captured to obtain dangerous images; The dangerous image is sent to the user terminal to complete the output.

2. The image data acquisition method under dangerous conditions according to claim 1, characterized in that, Preprocessing the original image to obtain the initial image includes: The first convolution tensor is obtained by performing a 3×3 convolution operation on the original image; The first convolutional tensor is subjected to average pooling and max pooling respectively to obtain the first average tensor and the first max tensor; The first average tensor and the first maximum tensor are concatenated to obtain the concatenated tensor; Substituting the spliced ​​vector into a multilayer perceptron yields a channel feature map; The first feature map is obtained by multiplying the monitoring image and the channel feature map; After performing a 3×3 convolution operation on the first feature map, a sigmoid activation function is input to obtain a spatial feature map. The initial image is obtained by multiplying the channel feature map, spatial feature map, and first feature map.

3. The image data acquisition method under dangerous conditions according to claim 1, characterized in that, The principle and process of the hazardous scene identification model include: The initial image is processed sequentially through a first convolutional layer, a second convolutional layer, a first star layer, a third convolutional layer, and a second star layer to obtain a first feature map; The first feature map is processed sequentially through the fourth convolutional layer and the third star layer to obtain the second feature map; The second feature map is processed sequentially through the fifth convolutional layer, the fourth star layer, the feature enhancement layer, and the SPPF layer to obtain the third feature map; The third feature map is upsampled and then fused with the second feature map to obtain the fourth feature map. The fourth feature map is processed through the C2f_ML layer to obtain the fifth feature map; the fifth feature map is then upsampled and fused with the first feature map to obtain the sixth feature map; The sixth feature map is processed through a C2f_ML layer and a sixth convolutional layer to obtain the seventh feature map; the seventh feature map and the fifth feature map are then fused to obtain the eighth feature map; The eighth feature map is processed through a C2f_ML layer and a seventh convolutional layer to obtain the ninth feature map; the ninth feature map and the third feature map are then fused to obtain the ninth feature map. The ninth feature map is processed through the C2f_ML layer to obtain the target feature map; The target feature map is used to search a preset database to determine the identification feature value.

4. The image data acquisition method under dangerous conditions according to claim 3, characterized in that, The working principle of the star-shaped layer includes: Use the feature map of the input star layer as the input feature; The input features are processed through the DWConv layer to obtain the first feature; Substituting the first feature into the first fully connected layer and the second fully connected layer in sequence, we obtain the first fully connected feature and the second fully connected feature, respectively. The second feature is obtained by multiplying the first fully connected feature and the second fully connected layer feature; The second feature is substituted into the third fully connected layer to obtain the third feature; The third feature is substituted into the DWConv layer for processing to obtain the target feature.

5. The image data acquisition method under dangerous conditions according to claim 3, characterized in that, The working principle of the C2f_ML layer includes: Use the feature map of the input C2f_ML layer as the input feature tensor; The first feature tensor is obtained by substituting the input feature tensor into the Split layer; The first feature tensor is substituted into the first Bottleneck layer to obtain the second feature tensor; the second feature tensor is substituted into the second Bottleneck layer to obtain the third feature tensor. Substitute the first feature tensor, the second feature tensor, and the third feature tensor into the Concat layer to obtain the fused feature tensor; The fused feature tensor is substituted into the convolutional layer to obtain the output feature tensor.

6. The image data acquisition method under dangerous conditions according to claim 5, characterized in that, The working principle of the Bottleneck layer includes: The Bottleneck layer is composed of two convolutional layers and an ML layer connected in sequence. The specific working principle of the ML layer includes: The feature map input to the ML layer is processed by a local attention pooling layer to obtain attention pooling features; The attention pooling features are obtained by performing a 1×1 convolution operation on the attention pooling features; The attention pooling features and the attention pooling features are fused to obtain the fused features; The local target features are obtained by multiplying the feature map and the fused features.

7. The method for acquiring image data under dangerous conditions according to claim 3, characterized in that, The principle and process of the feature enhancement layer include: Where Y1 is the input of the feature enhancement layer, Y2, Y3, Y4 and Y5 are the outputs of the four two-dimensional convolution branches respectively; Z1 is the attention weight; Z is the output of the feature enhancement module; f is the operator, the symbol ⊗ indicates element-wise multiplication, the superscript indicates the kernel size, and the subscript Conv2d indicates two-dimensional convolution.

8. The method for acquiring image data under dangerous conditions according to claim 1, characterized in that, The step of determining scene hazard information from the monitored image based on the identified feature values ​​includes: If the identified feature value is greater than the preset feature threshold, the scene of the detected image is in a dangerous state.

9. The method for acquiring image data under dangerous conditions according to claim 8, characterized in that, The method of determining scene hazard information in the monitored image based on the identified feature values ​​also includes: If the identified feature value is less than or equal to the preset feature threshold, the scene of the detected image is in a safe state.

Images