Polarization imaging earth surface unexploded ordnance identification device and method based on deep learning

Abstract

The invention discloses a polarization imaging earth surface unexploded ordnance identification device and method based on deep learning, the device comprises a light source module, a detection module and a computer, the light source module and the detection module are both wirelessly connected with the computer, and the light source module comprises a first unmanned aerial vehicle and a polarized light illumination assembly installed on the first unmanned aerial vehicle; the detection module comprises a second unmanned aerial vehicle, a holder carried by the second unmanned aerial vehicle and a positioning system, and a polarization image acquisition device is mounted on the holder; the computer controls the flight height of the first unmanned aerial vehicle and the second unmanned aerial vehicle and the rotation angle of the holder, the polarized light illumination assembly emits polarized light to irradiate a detection area, the polarized image acquisition device shoots the detection area to obtain a polarized image, polarized image data is input into the computer for model training and optimization based on deep learning, and the detection area is obtained. Therefore, the unexploded ordnance in the polarization image is detected and identified. The unmanned aerial vehicle is utilized, deep learning and polarization technologies are combined, and the earth surface unexploded ordnance can be automatically and accurately detected in an all-weather mode.

Description

Technical Field

[0001] The present invention relates to the field of detection and identification technology, and in particular to a device and method for identifying unexploded bombs on the surface using polarization imaging based on deep learning. Background Art

[0002] At present, there are still many countries and regions in the world that are in the midst of war, and inevitably there are a large number of unexploded mines, artillery shells, missiles, and various submunitions. Unexploded bombs are usually scattered near the surface, posing a great threat to the property and safety of local residents. Even in places where peace has been restored, unexploded weapons left over from historical wars still pose long-term safety risks. Developing precise detection technology to locate these potential threats is of great significance to protecting people's lives and property.

[0003] With the development of unexploded weapon systems such as aerial bombs and cluster bombs, the detection technology of unexploded ordnance on the surface is also improving. Common detection methods include physical detection and biological detection. Physical detection is low-cost but easily affected by environmental interference and has low accuracy; biological detection has a wide coverage and is reliable, but it takes a long time and has low efficiency. Intelligent detection technology performs well in data processing and mainly identifies target features by analyzing large amounts of data. By using deep learning algorithms to continuously optimize existing data, target templates can be created and matched with actual targets, improving recognition speed and accuracy.

[0004] The identification of unexploded ordnance on the surface mainly relies on digital imaging technology, including visible light, infrared and polarization imaging. These technologies can display information such as the shape, size and color of the target. Visible light imaging is fast, but it is difficult to distinguish when the target and the background color are similar. Infrared imaging can effectively distinguish metal objects from the background, but it is not easy to capture details and is easily affected by the environment. Polarization imaging can not only provide shape and color information, but also obtain characteristics such as phase and intensity, and suppress interference from the atmosphere, haze and background noise.

[0005] Polarization imaging target recognition based on deep learning is a highly promising recognition method. It combines polarization imaging with deep learning to perform target recognition in complex environments and further improve recognition accuracy. Summary of the invention

[0006] In order to solve the above technical problems existing in the prior art, the present invention proposes a device and method for identifying unexploded bombs on the ground using polarization imaging based on deep learning, and the specific technical solution is as follows: A surface unexploded bomb polarization imaging recognition device based on deep learning comprises a light source module, a detection module and a computer, wherein the light source module and the detection module are both wirelessly connected to the computer; the light source module comprises: a first unmanned aerial vehicle, on which a polarized light illumination component is arranged and installed; the detection module comprises: a second unmanned aerial vehicle, on which a gimbal and a positioning system are mounted, and a polarized image acquisition device is installed on the gimbal; a control system, an information processing system and a training system are installed in the computer; the control system is used to control the flying height of the first unmanned aerial vehicle and the second unmanned aerial vehicle and the rotation angle of the gimbal, the polarized light illumination component is used to emit polarized light to illuminate a detection area, the polarized image acquisition device shoots the illuminated detection area after the angle is controlled by the gimbal to obtain a polarized image, the positioning system is used to obtain the position information of the second unmanned aerial vehicle, the information processing system is used to receive the position information and polarized image data of the second unmanned aerial vehicle, and the training system is used to input the polarized image data for model training optimization, so as to realize the detection of unexploded bombs in the polarized image.

[0007] Furthermore, the polarized light illumination assembly comprises: a light source, a polarizer, a beam expander lens group, and an optical window. The light source, the polarizer, the beam expander lens group, and the optical window are installed in sequence on the same optical axis and arranged longitudinally.

[0008] Furthermore, the positioning system adopts the Beidou positioning system.

[0009] Furthermore, the polarization image acquisition device adopts a full-Stokes polarization camera.

[0010] A method for identifying surface unexploded bombs by polarization imaging based on deep learning, comprising the following steps: Step 1: Select a detection area at a location suspected of containing unexploded bombs, operate a first drone to hover directly above the detection area, and use a polarized light illumination component to emit polarized light to illuminate the detection area; Step 2: Plan a semicircular flight trajectory with a pitch angle of 180°, taking the distance from the light source module to the detection area as the radius; Step 3: Control the second UAV to fly at a constant speed around the pitch semicircular arc trajectory planned in step 2, and take polarization images at the same pitch angle intervals; Step 4: hover the second drone on the same plane as the first drone, and plan a 180° semicircular flight trajectory with a radius of the distance from the light source module to the detection area; Step 5: Control the second UAV to fly at a constant speed around the azimuth semicircular arc trajectory planned in step 4, and take polarization images at the same azimuth angle intervals; Step 6: Transmit the polarization image data captured in Step 3 and Step 5 to a computer information processing system, and send the location information of the second UAV; Step 7: Control the first UAV and the second UAV to move according to a preset distance, and repeat steps 1 to 6 to detect the next detection area; Step 8: First, manually determine whether there is an unexploded bomb in the previous area. After verification, the polarization images containing the unexploded bombs at various angles are used as a data set and input into the computer training system for training; Step 9. After the training is completed, the images obtained in step 3 and step 5 are input into the computer for detection, and the computer is used to directly determine whether there is an unexploded bomb in the polarization image; at the same time, the image data of the unexploded bomb detected and identified is expanded to the training data set and continued to be input into the training system for training.

[0011] Furthermore, in step three, a polarization image is taken for every 5° pitch angle of flight, for a total of 37 polarization images detected at different pitch angles; in step five, a polarization image is taken for every 5° azimuth angle of flight, for a total of 37 polarization images detected at different azimuth angles.

[0012] Furthermore, the training system adopts the YOLOv8 model as the target detection model, and a channel prior convolutional attention mechanism is added to the YOLOv8 model.

[0013] Furthermore, the detection process of the YOLOv8 model includes: a) Resize the input image to the size required by the model and normalize it; b) Perform a forward pass on the input image to extract features and generate predictions: Bounding box regression: predict the bounding box position (x, y, w, h) of each object in the image, where x and y are the coordinates of the center point of the bounding box, w is the width of the image, and h is the height of the image; Category prediction: predict the category of each target; Confidence score: Output the confidence of each bounding box to indicate the probability of the object existing in the bounding box; c) For each category, apply the non-maximum suppression algorithm to control the degree of overlap by setting a threshold, thereby removing overlapping bounding boxes and retaining only the box with the highest confidence; d) Return the final bounding box, category, and confidence, and display them visually on the original image.

[0014] Through the above design scheme, the present invention can bring the following beneficial effects: 1. Using drones, unexploded bombs on the ground can be detected and identified automatically around the clock.

[0015] 2. Combining deep learning with polarization technology has increased the technical means of identifying unexploded bombs on the surface. It can not only find the characteristic information of the target from a large amount of data, but also use the identification data to continuously expand the training data set and continuously improve the recognition accuracy.

[0016] 3. Achieve accurate detection and positioning of unexploded bomb targets, ensure personnel safety and equipment mobility, and enhance national defense science and technology capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a structural principle diagram of a surface unexploded bomb polarization imaging identification device based on deep learning according to an embodiment of the present invention; Figure 2 Schematic diagram of the improved yolov8 model structure of an embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical scheme and technical effect of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0019] like Figure 1 As shown, a surface unexploded bomb polarization imaging identification device based on deep learning of the present invention includes a light source module, a detection module and a computer, and the light source module and the detection module are both wirelessly connected to the computer. The light source module includes: a first drone, a light source, a polarizer, a beam expander, and an optical window; the detection module includes: a second drone, a polarization image acquisition device, and a gimbal; the computer is installed with an information processing system, a control system, and a training system.

[0020] The light source, polarizer, beam expander group, and optical window are mounted in a first unmanned aerial vehicle. The light source, polarizer, beam expander group, and optical window are installed in sequence on the same optical axis and arranged longitudinally. The polarizer is used for polarization. The beam expander group is used for expanding the light beam and then emitting it through the optical window to achieve large field of view illumination.

[0021] The gimbal is installed on the second UAV, and the polarization image acquisition device is installed on the gimbal, and the gimbal controls the angle adjustment of the polarization image acquisition device. The polarization image acquisition device uses a full-Stokes polarization camera to capture polarization images of the surface position illuminated by polarized light. The second UAV is also equipped with a Beidou positioning system, which acquires and sends the position information of the second UAV to a computer in real time.

[0022] The control system is used to control the flying height of the first UAV and the second UAV and the rotation angle of the gimbal; the information processing system receives the position information and polarization image data of the second UAV; the training system is used to input the polarization image data for model training optimization to realize the detection of unexploded bombs in the polarization image.

[0023] Based on the above identification device, the present invention also provides a surface unexploded bomb polarization imaging identification method based on deep learning, comprising the following steps: Step 1: Select a detection area at a location suspected of containing unexploded bombs, operate the first drone to hover directly above the detection area, and emit polarized light to illuminate the detection area; Step 2: Plan a semicircular flight trajectory with a pitch angle of 180°, taking the distance from the light source module to the detection area as the radius; Step 3: Control the second UAV to fly at a constant speed around the pitch semicircular arc trajectory planned in step 2, and take a polarization image every 5° of flight, for a total of 37 polarization images detected at different pitch angles; Step 4: hover the second drone on the same plane as the first drone, and plan a 180° semicircular flight trajectory with a radius of the distance from the light source module to the detection area; Step 5: Control the second UAV to fly at a constant speed around the azimuth semicircular arc trajectory, and take a polarization image every 5° of flight, for a total of 37 polarization images detected at different azimuth angles; Step 6: Transmit the 74 polarization image data taken to the computer information processing system, and send the location information of the second UAV; Step 7: Control the first UAV and the second UAV to move according to a preset distance, and repeat steps 1 to 6 to detect the next detection area; Step 8: First, manually determine whether there is an unexploded bomb in the previous area. After verification, the polarization images containing the unexploded bombs at various angles are used as training data sets and input into the computer training system for training; Step 9. After the training is completed, the images obtained in step 3 and step 5 are input into the computer for detection, and the computer is used to directly determine whether there is an unexploded bomb in the polarization image; at the same time, the image data of the unexploded bomb detected and identified is expanded to the training data set, and continues to be input into the training system for training to continuously improve the recognition accuracy.

[0024] Specifically, the data set training method process in step eight is as follows: 1. Use drones to photograph unexploded bombs, perform manual annotation, and collect data sets.

[0025] 2. The target detection model used by the training system is the YOLOv8 model. The YOLOv8 model is optimized according to the characteristics of the UAV detection of unexploded bombs: the channel prior convolutional attention (CPCA) mechanism is added to enhance feature representation and dynamically allocate attention weights to improve the detection and recognition accuracy of small targets. Figure 2 shown.

[0026] 3. After training is completed, target detection can be performed. The detection process of the YOLOv8 model is as follows: a) Resize the input image to the size required by the model, such as 640x640, and normalize it.

[0027] b) The input image is forward propagated through the YOLOv8 model to extract features and generate predictions: Bounding box regression: The model predicts the bounding box position (x, y, w, h) of each object, where x and y are the coordinates of the center point of the bounding box, w is the width of the image, and h is the height of the image; Category prediction: The model predicts the category of each target; Confidence score: The model outputs the confidence score of each bounding box, which indicates the probability that an object exists in the box.

[0028] c) For each category, apply the non-maximum suppression (NMS) algorithm to remove overlapping bounding boxes and keep only the boxes with the highest confidence. The degree of overlap is controlled by setting a threshold; d) Return the final bounding box, category, and confidence, and display them visually on the original image.

[0029] 4. The initial data set is collected by drones and manually labeled. After achieving basic recognition effects, a combination of artificial intelligence and manual methods is used to label and collect data sets, enrich the data sets, and iterate the algorithm to achieve better recognition effects.

[0030] The present invention combines polarization imaging with deep learning, and can directly determine whether an area contains unexploded bombs through a computer system, thereby achieving efficient identification of surface unexploded bombs in complex environments.

[0031] The above is only a preferred implementation case of the present invention and does not limit the present invention in any form. Although the implementation process of the present invention is described in detail above, for those familiar with the art, they can still modify the technical solutions recorded in the above examples, or replace some of the technical features therein with equivalents. All modifications, equivalent replacements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A surface unexploded bomb polarization imaging identification device based on deep learning, comprising a light source module, a detection module and a computer, wherein the light source module and the detection module are both wirelessly connected to the computer; It is characterized in that The light source module comprises: a first drone, on which a polarized light illumination component is installed; The detection module includes: a second drone, the second drone is equipped with a gimbal and a positioning system, and a polarization image acquisition device is installed on the gimbal; The computer is installed with a control system, an information processing system, and a training system; The control system is used to control the flying height of the first UAV and the second UAV and the rotation angle of the gimbal. The polarized light illumination component is used to emit polarized light to illuminate the detection area. The polarized image acquisition device shoots the illuminated detection area after the angle is controlled by the gimbal to obtain a polarized image. The positioning system is used to obtain the position information of the second UAV. The information processing system is used to receive the position information and polarized image data of the second UAV. The training system is used to input the polarized image data for model training and optimization, so as to realize the detection of unexploded bombs in the polarized image.

2. The surface unexploded bomb polarization imaging identification device based on deep learning according to claim 1 is characterized in that: The polarized light illumination assembly comprises: a light source, a polarizer, a beam expander lens group, and an optical window. The light source, the polarizer, the beam expander lens group, and the optical window are installed in sequence on the same optical axis and arranged longitudinally.

3. The surface unexploded bomb polarization imaging identification device based on deep learning according to claim 1 is characterized in that: The positioning system adopts the Beidou positioning system.

4. The surface unexploded bomb polarization imaging identification device based on deep learning according to claim 1 is characterized in that: The polarization image acquisition device adopts a full-Stokes polarization camera.

5. A method for identifying unexploded bombs on the ground using polarization imaging of unexploded bombs on the ground based on deep learning according to any one of claims 1 to 4, characterized in that: The following steps are involved: Step 1: Select a detection area at a location suspected of containing unexploded bombs, operate a first drone to hover directly above the detection area, and use a polarized light illumination component to emit polarized light to illuminate the detection area; Step 2: Plan a semicircular flight trajectory with a pitch angle of 180°, taking the distance from the light source module to the detection area as the radius; Step 3: Control the second UAV to fly at a constant speed around the pitch semicircular arc trajectory planned in step 2, and take polarization images at the same pitch angle intervals; Step 4: hover the second drone on the same plane as the first drone, and plan a 180° semicircular flight trajectory with a radius of the distance from the light source module to the detection area; Step 5: Control the second UAV to fly at a constant speed around the azimuth semicircular arc trajectory planned in step 4, and take polarization images at the same azimuth angle intervals; Step 6: Transmit the polarization image data captured in Step 3 and Step 5 to a computer information processing system, and send the location information of the second UAV; Step 7: Control the first UAV and the second UAV to move according to a preset distance, and repeat steps 1 to 6 to detect the next detection area; Step 8: First, manually determine whether there is an unexploded bomb in the previous area. After verification, the polarization images containing the unexploded bombs at various angles are used as training data sets and input into the computer training system for training; Step 9. After the training is completed, the images obtained in step 3 and step 5 are input into the computer for detection, and the computer is used to directly determine whether there is an unexploded bomb in the polarization image; at the same time, the image data of the unexploded bomb detected and identified is expanded to the training data set and continued to be input into the training system for training.

6. The method for identifying unexploded bombs on the ground by polarization imaging according to claim 5, characterized in that: In step three, a polarization image is taken for every 5° pitch angle of flight, for a total of 37 polarization images detected at different pitch angles; in step five, a polarization image is taken for every 5° azimuth angle of flight, for a total of 37 polarization images detected at different azimuth angles.

7. The method for identifying unexploded bombs on the ground by polarization imaging according to claim 5, characterized in that: The training system adopts the YOLOv8 model as the target detection model, and a channel prior convolutional attention mechanism is added to the YOLOv8 model.

8. The method for identifying unexploded bombs on the ground by polarization imaging according to claim 7, characterized in that: The detection process of the YOLOv8 model includes: a) Resize the input image to the size required by the model and normalize it; b) Perform a forward pass on the input image to extract features and generate predictions: Bounding box regression: predict the bounding box position (x, y, w, h) of each object in the image, where x and y are the coordinates of the center point of the bounding box, w is the width of the image, and h is the height of the image; Category prediction: predict the category of each target; Confidence score: Output the confidence of each bounding box to indicate the probability of the object existing in the bounding box; c) For each category, apply the non-maximum suppression algorithm to control the degree of overlap by setting a threshold, thereby removing overlapping bounding boxes and retaining only the box with the highest confidence; d) Return the final bounding box, category, and confidence, and display them visually on the original image.

Images