An AI Vision-based Dynamic Image Recognition System

Abstract

The present application provides an AI vision-based dynamic image recognition system, which is applied to the field of image processing technology. The visible light image and the infrared image are synchronously acquired through the image acquisition module, and multi-modal image information is utilized. The blur degree evaluation module obtains the blur degree parameter β. The enhancement coefficient calculation module adaptively calculates the edge enhancement coefficient α according to the β value. The edge enhancement module detects the edges of the infrared image and enhances the edge information. The image fusion module performs pixel-level fusion on the enhanced infrared image and the visible light image to obtain a fused image. The target recognition module performs target detection and recognition on the fused image, obtains the position and category information of the target, and provides shooting feedback to the user based on this, thereby completing the virtual-real interactive shooting training. Therefore, the present application has the beneficial effect of overcoming the negative impact of light-absorbing materials in low-light environments and improving the target recognition accuracy of multi-modal image fusion.

Description

Technical Field

[0001] This application relates to the field of computer technology, and particularly to an AI vision-based dynamic image recognition system. Background Art

[0002] In an indoor shooting training venue, trainees usually use an AI vision-based dynamic image recognition system for moving target shooting training. To more realistically simulate the target characteristics in a battlefield environment and reduce visible light reflection, the target is often a human silhouette target made of light-absorbing material. At the same time, to simulate an energy-saving environment or a specific combat scenario, the training venue usually uses low-light illumination. In order to effectively identify the target even under low-light conditions, the AI vision-based dynamic image recognition system integrates multi-modal image recognition technology, that is, fuses visible light and infrared images for target detection and recognition. To meet the low-latency requirements of shooting training, image recognition processing is usually performed on edge computing devices on the shooting range side. However, when a low-light environment and a light-absorbing material target coexist, the light-absorbing material absorbs infrared light, resulting in a blurred target contour in the infrared image, directly affecting the effect of multi-modal image fusion, and further reducing the accuracy of target recognition. Existing image recognition technologies are difficult to effectively solve the problem of blurred infrared images caused by low light and light-absorbing material targets under the condition of limited edge computing resources, resulting in limited performance of the virtual-real interactive shooting training system.

[0003] Therefore, in complex scenarios such as indoor low-light and light-absorbing material human silhouette moving target shooting training, how to effectively overcome the problem of blurred target contours in infrared images caused by light-absorbing materials absorbing infrared light, and under the low-latency constraint of edge computing, achieve effective enhancement of blurred infrared images, and further improve the target recognition accuracy of multi-modal image fusion to ensure the accuracy and real-time performance of virtual-real interactive shooting training has become an urgent technical problem to be solved. Summary of the Invention

[0004] In view of the above deficiencies of the prior art, this application provides an AI vision-based dynamic image recognition system, which is applied to the technical field and has the advantage of improving the target recognition accuracy in a low-light and light-absorbing target scenario.

[0005] In a first aspect, an AI vision-based dynamic image recognition system for virtual-real interactive shooting training of light-absorbing targets in an indoor low-light environment, the system includes:

[0006] An image acquisition module: used to obtain visible light images and infrared images containing human silhouette targets synchronously collected by a visible light camera and an infrared camera;

[0007] A blur degree evaluation module: used to evaluate the blur degree of the infrared image and obtain a blur degree parameter β;

[0008] Enhancement coefficient calculation module: used to calculate the adaptive edge enhancement coefficient α according to the blur degree parameter β;

[0009] Edge enhancement module: used to perform edge detection on the infrared image using the Laplacian operator, obtain the infrared image edge information, enhance the infrared image edge information according to the adaptive edge enhancement coefficient α, and obtain the infrared image after edge enhancement;

[0010] Image fusion module: used to perform pixel-level fusion on the infrared image after edge enhancement and the visible light image to obtain a fused image;

[0011] Target recognition module: used to perform target detection and recognition on the fused image, obtain the target position and category information of the human silhouette target, and provide shooting feedback to the user according to the target position and the category information.

[0012] An AI vision-based dynamic image recognition system provided by the present application aims to solve the problem of reduced target recognition accuracy in virtual-real interaction shooting training due to blurred infrared images under indoor low-light and light-absorbing target conditions. The system first synchronously acquires visible light images and infrared images through an image acquisition module, utilizing multi-modal image information. Aiming at the possible blur problem of infrared images, a blur degree evaluation module is designed to quantify the blur degree of infrared images and obtain the blur degree parameter β. The enhancement coefficient calculation module then adaptively calculates the edge enhancement coefficient α according to the β value, realizing the linkage adjustment of the blur degree and the enhancement intensity. The edge enhancement module uses the Laplacian operator to detect the edges of the infrared image and enhances the edge information according to the adaptive edge enhancement coefficient α, thereby improving the clarity of the blurred infrared image. The image fusion module performs pixel-level fusion on the enhanced infrared image and the visible light image, comprehensively utilizing the advantages of the two modal images to obtain a fused image with richer information. Finally, the target recognition module performs target detection and recognition on the fused image, obtains the position and category information of the target, and provides shooting feedback to the user accordingly, completing the virtual-real interaction shooting training. Therefore, the present application has the beneficial effect of overcoming the negative impact of light-absorbing materials in low-light environments and effectively enhancing blurred infrared images under the low-latency constraint of edge computing, improving the target recognition accuracy of multi-modal image fusion to ensure the accuracy and real-time performance of virtual-real interaction shooting training.

[0013] Further, the blur degree evaluation module includes:

[0014] Gradient calculation module: used to calculate the image gradient values of multiple frames of the infrared image;

[0015] Image sharpness evaluation module: used to calculate an image sharpness evaluation index according to multiple said image gradient values: σ = ∑(G - μ)^2 / N, where σ is the said image sharpness evaluation index, G is the said image gradient value, μ is the average image gradient, and N is the total number of image pixels;

[0016] Blur parameter calculation module: used to calculate the mean value of the image sharpness evaluation index σ of multiple frames of infrared images and use this mean value as the blur degree parameter β.

[0017] A dynamic image recognition system based on AI vision provided by the present application, through the collaborative work of a gradient calculation module, an image sharpness evaluation module, and a blur parameter calculation module, provides an effective method for quantifying the blur degree of infrared images, and provides a reliable basis for blur degree parameters for subsequent adaptive image enhancement and image fusion.

[0018] Further, the enhancement coefficient calculation module includes:

[0019] Preset coefficient acquisition module: used to acquire a preset minimum enhancement coefficient α_min, a maximum enhancement coefficient α_max, and a maximum blur degree index threshold β_max;

[0020] Enhancement coefficient determination module: used to adopt the maximum enhancement coefficient α_max as the adaptive edge enhancement coefficient α when the difference between the blur degree parameter β and β_max is less than a first preset difference;

[0021] When the difference between the blur degree parameter β and 0 is less than the first preset difference, adopt the minimum enhancement coefficient α_min as the adaptive edge enhancement coefficient α;

[0022] When the differences between the blur degree parameter β and 0 and β_max are both greater than or equal to the first preset difference, the adaptive edge enhancement coefficient α is calculated by the formula α = α_min + (α_max - α_min) * (β / β_max).

[0023] A dynamic image recognition system based on AI vision provided by the present application, through a piecewise linear adaptive adjustment strategy, enables the edge enhancement coefficient to be finely adjusted according to the actual blur degree of the infrared image, ensuring that noise is not introduced due to over-enhancement when the infrared image is clear, and effective enhancement is performed when the infrared image is blurred to improve the image quality, thereby improving the environmental adaptability and robustness of the system, solving the problem of blurred infrared images caused by low light and light-absorbing material targets mentioned in the background technology, and providing higher-quality infrared image edge information for subsequent image fusion and target recognition.

[0024] Further, the edge enhancement module includes:

[0025] Noise reduction processing module: used to perform Gaussian filtering on the infrared image, convolve with the Gaussian kernel and the infrared image to obtain the denoised infrared image I_IR_filtered;

[0026] Operator calculation module: calculate the adaptive Laplacian operator according to the blurriness parameter β;

[0027] Edge detection module: used to perform edge detection on the denoised infrared image I_IR_filtered using the Laplacian operator to obtain the edge information of the infrared image;

[0028] Edge enhancement sub-module: used to fuse the edge information of the infrared image and the original infrared image according to the adaptive edge enhancement coefficient α to obtain the edge-enhanced infrared image I_IR_enhanced.

[0029] An AI vision-based dynamic image recognition system provided by the present application. The edge enhancement module first performs Gaussian filtering on the infrared image through the noise reduction processing module to reduce image noise and provide higher-quality image data for subsequent edge detection. Then, the operator calculation module no longer uses a fixed Laplacian operator, but adaptively calculates the Laplacian operator according to the blurriness parameter β, enabling the edge detection operator to be adjusted according to the image blurriness and more accurately detect the edge information of the blurred image. Next, the edge detection module uses the adaptive Laplacian operator to perform edge detection on the denoised infrared image to obtain more accurate edge information of the infrared image. Finally, the edge enhancement sub-module fuses the extracted edge information of the infrared image with the original infrared image according to the adaptive edge enhancement coefficient α to achieve adaptive enhancement of the edges of the infrared image. Through the collaborative work of the above modules, the edge enhancement module provided by this solution can adaptively adjust the edge enhancement processing process according to the blurriness of the infrared image, effectively improve the edge enhancement effect of the blurred infrared image, provide higher-quality image data for subsequent image fusion and target recognition, and thus improve the target recognition accuracy and system performance of the virtual-real interaction shooting training system.

[0030] Further, the operator calculation module includes:

[0031] Mapping establishment module: used to pre-establish a mapping table between the blurriness parameter β and the linear mapping parameter γ, and query the mapping table according to the blurriness parameter β to obtain the linear mapping parameter γ;

[0032] Operator calculation sub-module: It is used for the minimum parameter L_min and the maximum parameter L_max of the preset Laplace operator. According to the linear mapping parameter γ, the Laplace operator parameter L is calculated through the formula L = L_min + (L_max - L_min) * γ.

[0033] Furthermore, the image fusion module includes:

[0034] Normalization processing module: It is used to perform normalization processing on the edge-enhanced infrared image and the visible light image to obtain a normalized infrared image and a normalized visible light image;

[0035] Information entropy calculation module: It is used to perform image segmentation on the normalized infrared image and the normalized visible light image respectively to obtain a normalized infrared image block and a normalized visible light image block, and calculate the information entropy of each image block;

[0036] Fusion image sub-module: It is used to fuse the normalized infrared image block and the normalized visible light image block in a weighted fusion manner according to the information entropy of each image block to obtain the fusion image.

[0037] Furthermore, the fusion image sub-module includes:

[0038] Weight mapping table establishment module: It is used to pre-statistically calculate the information entropy range of each image block in the training image set and establish a weight mapping table between the information entropy range and the weight coefficient;

[0039] Initial weight acquisition module: It is used to query and obtain the initial weight coefficients of the normalized infrared image block and the normalized visible light image block according to the weight mapping table;

[0040] Image block gradient value calculation module: It is used to calculate the image block gradient values of the normalized infrared image block and the normalized visible light image block;

[0041] Weight coefficient adjustment module: It is used to adjust the initial weight coefficients of each image block according to the image block gradient value to obtain the adjusted weight coefficients of the infrared image block and the visible light image block respectively;

[0042] First fusion image sub-module: It is used to fuse the normalized infrared image block and the normalized visible light image block according to the adjusted weight coefficients of the infrared image block and the visible light image block to obtain the fusion image block, and then combine all the fusion image blocks into the complete fusion image.

[0043] Furthermore, the weight coefficient adjustment module includes:

[0044] Weight calculation module: used to calculate the adjusted weight coefficients W_IR_adjusted of the infrared image block and W_VIS_adjusted of the visible light image block through the formulas W_IR_adjusted = W_IR_initial + k * G_IR and W_VIS_adjusted = W_VIS_initial + k * G_VIS; where W_IR_initial and W_VIS_initial are the initial weight coefficients of the infrared and visible light image blocks respectively, G_IR and G_VIS are the gradient values of the infrared and visible light image blocks respectively, and k is a preset gradient weight coefficient adjustment factor.

[0045] Further, the target recognition module includes:

[0046] Target information acquisition module: used to process the fused image with a lightweight YOLOv5 target detection network to obtain the initial position information and category information of the human silhouette target;

[0047] Prediction module: used to track the human silhouette target using a motion target tracking algorithm based on a Kalman filter according to the initial position information and the category information, predict the position information of the human silhouette target in the next frame of image, and smooth the predicted position information to obtain accurate target position information;

[0048] Feedback module: used to generate a three-dimensional ballistic trajectory according to the accurate target position information and the category information, and add the three-dimensional ballistic trajectory to the shooting feedback to feedback to the user.

[0049] Further, the feedback module includes:

[0050] Ballistic trajectory generation module: used to calculate the three-dimensional space vector from the shooting device to the target according to the accurate target position information and the category information, combined with the virtual-real interaction system, simulate the bullet flight trajectory according to the ballistic model, and generate a three-dimensional ballistic trajectory;

[0051] Shooting effect generation module: used to judge whether the target is hit according to the generated three-dimensional ballistic trajectory; if it is determined that the target is hit, record the hit position information and the shooting force, and perform damage modeling on the three-dimensional model of the target according to the hit position information and the shooting force to generate a damaged effect;

[0052] Feedback sub-module: used to add the damaged effect to the shooting feedback to feedback to the user.

[0053] Beneficial effects: An AI vision-based dynamic image recognition system proposed in this application aims to solve the problem of reduced target recognition accuracy in virtual-real interactive shooting training due to blurred infrared images under indoor low-light and light-absorbing target conditions. The system first synchronously acquires visible light images and infrared images through an image acquisition module, leveraging multi-modal image information. To address the possible blur problem of infrared images, a blur degree evaluation module is designed to quantify the blur degree of infrared images, obtaining a blur degree parameter β. The enhancement coefficient calculation module then adaptively calculates an edge enhancement coefficient α based on the β value, achieving a linkage adjustment between the blur degree and the enhancement intensity. The edge enhancement module uses the Laplacian operator to detect the edges of infrared images and enhances the edge information according to the adaptive edge enhancement coefficient α, thereby improving the clarity of blurred infrared images. The image fusion module performs pixel-level fusion of the enhanced infrared image and the visible light image, comprehensively utilizing the advantages of the two modal images to obtain a more information-rich fused image. Finally, the target recognition module performs target detection and recognition on the fused image, obtains the position and category information of the target, and provides shooting feedback to the user accordingly, completing the virtual-real interactive shooting training. Therefore, this application has the beneficial effect of overcoming the negative impact of light-absorbing materials in low-light environments and effectively enhancing blurred infrared images under the low-latency constraint of edge computing, improving the target recognition accuracy of multi-modal image fusion to ensure the accuracy and real-time performance of virtual-real interactive shooting training. Description of the Drawings

[0054] Figure 1 This is a structural diagram of an AI vision-based dynamic image recognition system proposed in this application.

[0055] Reference numeral description: 101, image acquisition module; 102, blur degree evaluation module; 103, enhancement coefficient calculation module; 104, edge enhancement module; 105, image fusion module; 106, target recognition module. Detailed Embodiments

[0056] Next, the technical solutions in the embodiments of this application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all of the embodiments. Usually, the components of the embodiments of this application described and marked here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the drawings is not intended to limit the scope of the claimed application, but merely represents the selected embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative efforts fall within the scope of protection of this application.

[0057] It should be noted that similar reference numerals and letters denote similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. At the same time, in the description of the present application, terms such as "first" and "second" are only used for distinguishing descriptions and cannot be construed as indicating or implying relative importance.

[0058] In complex scenarios such as indoor low-light and shooting training with a moving target of a human silhouette made of light-absorbing material, how to effectively overcome the problem of blurred infrared image target contours caused by the absorption of infrared light by light-absorbing materials, and under the low-latency constraint of edge computing, achieve effective enhancement of blurred infrared images, and then improve the target recognition accuracy of multi-modal image fusion to ensure the accuracy and real-time performance of virtual-real interactive shooting training has become a technical problem to be solved urgently. To solve this problem, the present application proposes an AI vision-based dynamic image recognition system as follows:

[0059] Please refer to Figure 1 , in the first aspect, an AI vision-based dynamic image recognition system for virtual-real interactive shooting training of light-absorbing targets in indoor low-light environments, the system includes:

[0060] An image acquisition module 101: for acquiring visible light images and infrared images containing human silhouette targets synchronously collected by a visible light camera and an infrared camera;

[0061] A blur degree evaluation module 102: for evaluating the blur degree of the infrared image to obtain a blur degree parameter β;

[0062] An enhancement coefficient calculation module 103: for calculating an adaptive edge enhancement coefficient α according to the blur degree parameter β;

[0063] An edge enhancement module 104: for performing edge detection on the infrared image using a Laplacian operator to obtain infrared image edge information, and enhancing the infrared image edge information according to the adaptive edge enhancement coefficient α to obtain an edge-enhanced infrared image;

[0064] An image fusion module 105: for performing pixel-level fusion on the edge-enhanced infrared image and the visible light image to obtain a fused image;

[0065] A target recognition module 106: for performing target detection and recognition on the fused image, obtaining the target position and category information of the human silhouette target, and providing shooting feedback to the user according to the target position and category information.

[0066] Among them, the image acquisition module 101 can use a visible light camera and an infrared camera to synchronously collect visible light images and infrared images containing human silhouette targets, thereby obtaining multi-modal image information and laying a data foundation for subsequent image processing.

[0067] The blur degree evaluation module 102 can calculate the image gradient values of multiple frames of infrared images, calculate the image sharpness evaluation index based on the image gradient values, and then use the mean value of the sharpness evaluation indexes of multiple frames of infrared images as the blur degree parameter β. Thus, the quality of infrared images can be quantified, providing a basis for adaptive image enhancement.

[0068] The enhancement coefficient calculation module 103 presets the minimum enhancement coefficient α_min, the maximum enhancement coefficient α_max, and the maximum blur degree index threshold β_max, and adaptively determines the edge enhancement coefficient α according to the relationship between the blur degree parameter β and β_max. Thus, the adaptability of image enhancement can be ensured, and the linkage adjustment of the blur degree and the enhancement intensity can be realized.

[0069] The edge enhancement module 104 can first perform Gaussian filtering and noise reduction processing on the infrared image, then use the Laplace operator for edge detection, and enhance the edge information according to the adaptive edge enhancement coefficient α. Thus, the edges of blurred infrared images can be effectively sharpened, highlighting the target contour.

[0070] The image fusion module 105 first normalizes the infrared image and the visible light image, and then fuses the normalized images in a weighted fusion manner. Thus, the information of both image modalities can be comprehensively utilized to improve the accuracy and robustness of target recognition.

[0071] The target recognition module 106 can use the lightweight YOLOv5 target detection network for target detection and the Kalman filter for target tracking. Thus, the function of virtual-real interactive shooting training can be realized, and the image processing results can be applied to actual training feedback.

[0072] Specifically, under the conditions of indoor low light and light-absorbing targets, this solution first synchronously acquires visible light images and infrared images through the image acquisition module 101. For the possible blurring problem of infrared images, the blurring degree evaluation module 102 quantifies the blurring degree of infrared images to obtain the blurring degree parameter β. The enhancement coefficient calculation module 103 adaptively calculates the edge enhancement coefficient α according to the β value, realizing the linkage adjustment of the blurring degree and the enhancement intensity. The edge enhancement module 104 uses the Laplace operator to detect the edges of infrared images and enhances the edge information based on the adaptive edge enhancement coefficient α to improve the clarity of blurred infrared images. The image fusion module 105 performs pixel-level fusion of the enhanced infrared images and visible light images, comprehensively utilizes the advantages of the two modal images, and obtains a more informative fused image. Finally, the target recognition module 106 performs target detection and recognition on the fused image, obtains the position and category information of the target, and provides shooting feedback for the user accordingly, completing the closed-loop of virtual-real interactive shooting training. Through the collaborative work of the above modules, the technical problem of reduced target recognition accuracy caused by blurred infrared images under indoor low light and light-absorbing target conditions is effectively solved, and the performance of the AI vision dynamic image recognition system is improved.

[0073] In some specific embodiments, the image acquisition module 101 uses a visible light camera with the model number DH-IPC-HFW5241EP-SE-S2 and an infrared camera with the model number DH-TPC-H3201 for image acquisition. The cameras are fixed on the edge computing device on the shooting range side to ensure that the field of view includes the human silhouette target. The blur degree evaluation module 102 first calculates the image gradient values of 5 consecutive infrared images, and then calculates the image sharpness evaluation index σ according to the formula σ = ∑(G - μ)^2 / N, where G is the image gradient value, μ is the average image gradient value, and N is the total number of image pixels. Finally, the average value of the image sharpness evaluation index σ of 5 images is calculated as the blur degree parameter β. The enhancement coefficient calculation module 103 presets the minimum enhancement coefficient α_min as 0.2, the maximum enhancement coefficient α_max as 0.8, and the maximum blur degree index threshold β_max as 0.05. When the difference between the blur degree parameter β and β_max is less than 0.01, the maximum enhancement coefficient α_max is used as the adaptive edge enhancement coefficient α; when the difference between the blur degree parameter β and 0 is less than 0.01, the minimum enhancement coefficient α_min is used as the adaptive edge enhancement coefficient α; in other cases, the adaptive edge enhancement coefficient α is calculated by the formula α = α_min + (α_max - α_min) * (β / β_max). The edge enhancement module 104 first performs Gaussian filtering on the infrared image using a Gaussian kernel with a size of 3x3 and a standard deviation of 0.5, then uses the Laplace operator for edge detection, and enhances the edge information according to the adaptive edge enhancement coefficient α. The image fusion module 105 uses a weighted fusion algorithm, and the weight coefficients are adaptively adjusted according to the information entropy and gradient values of the image blocks. The image block size is set to 8x8 pixels. The target recognition module 106 uses a lightweight YOLOv5s target detection network for target detection, and uses a Kalman filter to track the human silhouette target. The feedback module generates a three-dimensional ballistic trajectory and displays it in the virtual-real interaction interface, and at the same time generates a target breakage effect according to the hit position. Thus, the virtual-real interaction shooting training for the light-absorbing target in the indoor low-light environment can be realized, and the accuracy and real-time performance of target recognition can be ensured.

[0074] Further, the blur degree evaluation module 102 includes:

[0075] Gradient calculation module: used to calculate the image gradient values of multiple infrared images;

[0076] Image sharpness evaluation module: used to calculate the image sharpness evaluation index according to multiple image gradient values: σ = ∑(G - μ)^2 / N, where σ is the image sharpness evaluation index, G is the image gradient value, μ is the average image gradient value, and N is the total number of image pixels;

[0077] Blurriness parameter calculation module: used to calculate the mean value of the image sharpness evaluation index σ of multiple frames of infrared images, and use this mean value as the blurriness parameter β.

[0078] Among them, the gradient calculation module performs image gradient calculation operations on each frame of infrared image. The image gradient calculation operation can be implemented using a variety of existing technologies, such as the Sobel operator, Prewitt operator, or Canny operator, etc. These operators obtain the gradient value of the image at each pixel position by calculating the change rate of the gray value of the image pixels in the horizontal and vertical directions. The image gradient value characterizes the severity of the gray change in the local area of the image and reflects the edge and texture information of the image. The image sharpness evaluation module receives multiple image gradient values from the gradient calculation module. These image gradient values can be the gradient values of all pixel points from a single frame of infrared image, or a set of pixel point gradient values from multiple frames of infrared images.

[0079] The image sharpness evaluation module calculates the image sharpness evaluation index σ according to the received multiple image gradient values according to the formula σ = ∑(G - μ)^2 / N. Among them, G represents the image gradient value, μ represents the average image gradient value, and N represents the total number of image pixels. The average image gradient value μ is obtained by summing all image gradient values and dividing by the total number of image pixels N.

[0080] The blurriness parameter calculation module receives the image sharpness evaluation index σ of multiple frames of infrared images from the image sharpness evaluation module. The blurriness parameter calculation module calculates the mean value of these sharpness evaluation indexes σ. This mean value is used as the blurriness parameter β. The magnitude of the value of the blurriness parameter β reflects the blurriness of the infrared image. The higher the value, the clearer the image; the lower the value, the more blurred the image.

[0081] In some specific embodiments, the gradient calculation module uses the Sobel operator to calculate the image gradient value. When calculating the image sharpness evaluation index σ, the image sharpness evaluation module selects the gradient values of 5 consecutive frames of infrared images for calculation to improve the real-time performance and accuracy of the blurriness evaluation. The blurriness parameter calculation module calculates the arithmetic mean value of the sharpness evaluation indexes σ of these 5 frames of images and uses this mean value as the blurriness parameter β. The value range of the parameter β is limited between 0 and 100, where 0 represents a completely clear image and 100 represents a completely blurred image.

[0082] Furthermore, the enhancement coefficient calculation module 103 includes:

[0083] Preset coefficient acquisition module: used to obtain the preset minimum enhancement coefficient α_min, maximum enhancement coefficient α_max, and maximum blurriness index threshold β_max;

[0084] Enhancement coefficient determination module: When the difference between the fuzziness parameter β and β_max is less than the first preset difference, the maximum enhancement coefficient α_max is used as the adaptive edge enhancement coefficient α;

[0085] When the difference between the fuzziness parameter β and 0 is less than the first preset difference, the minimum enhancement coefficient α_min is used as the adaptive edge enhancement coefficient α;

[0086] When the differences between the fuzziness parameter β and 0 and between β and β_max are both greater than or equal to the first preset difference, the adaptive edge enhancement coefficient α is calculated by the formula α = α_min + (α_max - α_min) * (β / β_max).

[0087] Among them, the preset coefficient acquisition module is used to provide the parameter range for adjusting the edge enhancement coefficient and the fuzziness threshold. Specifically, the minimum enhancement coefficient α_min can be set to 0.1, the maximum enhancement coefficient α_max can be set to 1.0, and the maximum fuzziness index threshold β_max can be set to an empirical value, such as 100. The setting of these values provides a numerical basis for the subsequent calculation of the adaptive edge enhancement coefficient. The enhancement coefficient determination module is the core of this solution, which embodies the strategy of adaptively adjusting the enhancement coefficient. The first preset difference can be set to a small value, such as 5, to determine whether the fuzziness parameter β is close to the boundary value. When the difference between the fuzziness parameter β and β_max is less than 5, the system determines that the infrared image is very fuzzy. At this time, the maximum enhancement coefficient 1.0 is directly used to perform the maximum degree of edge enhancement, striving to extract more effective information from the fuzzy image. On the contrary, when the difference between the fuzziness parameter β and 0 is less than 5, the system determines that the infrared image is very clear. At this time, the minimum enhancement coefficient 0.1 is used for a small degree of edge enhancement, and α_min can even be set to 0 according to the actual situation to avoid introducing noise due to excessive enhancement of image clarity. In the intermediate state where the fuzziness parameter β is neither close to 0 nor close to β_max, the adaptive edge enhancement coefficient α is calculated by the linear interpolation formula. This linear calculation method ensures the smooth transition of the enhancement coefficient α with the fuzziness parameter β, realizing the adaptive adjustment of the edge enhancement degree. Thus, through the collaborative work of the preset coefficient acquisition module and the enhancement coefficient determination module, the accurate calculation of the adaptive edge enhancement coefficient is realized, ensuring that appropriate edge enhancement effects can be obtained under different fuzziness degrees.

[0088] Furthermore, the edge enhancement module 104 includes:

[0089] Noise reduction processing module: Used to perform Gaussian filtering on the infrared image, convolve the Gaussian kernel with the infrared image, and obtain the denoised infrared image I_IR_filtered;

[0090] Operator calculation module: Calculate the adaptive Laplacian operator according to the blur degree parameter β;

[0091] Edge detection module: Used to perform edge detection on the denoised infrared image I_IR_filtered using the Laplacian operator to obtain infrared image edge information;

[0092] Edge enhancement sub-module: Used to fuse the infrared image edge information with the original infrared image according to the adaptive edge enhancement coefficient α to obtain the edge-enhanced infrared image I_IR_enhanced.

[0093] Among them, the denoising processing module can use a Gaussian kernel with a size of 3x3 or 5x5 to filter the infrared image to reduce the interference of image noise on subsequent edge detection.

[0094] In the operator calculation module, the blur degree parameter β can be pre-divided into multiple levels. For example, the β value can be divided into multiple intervals such as [0, 20), [20, 40), [40, 60), [60, 80), [80, 100], etc. Each interval corresponds to a different value range of the linear mapping parameter γ. By querying the pre-established mapping table of the blur degree parameter β and the linear mapping parameter γ, the linear mapping parameter γ that matches the current blur degree parameter β can be quickly obtained. The Laplacian operator used in the edge detection module can be the standard 3x3 Laplacian operator template or other Laplacian operators of different sizes or forms. In the edge enhancement sub-module, the edge enhancement coefficient α can be adjusted according to the actual application scenario and requirements. For example, in the case of weak light and a high degree of target edge blur, the α value can be appropriately increased to improve the edge enhancement effect.

[0095] In some specific embodiments, the standard deviation σ value of the Gaussian kernel used for Gaussian filtering can be adjusted according to the actual noise level. For example, when the noise level is high, the σ value can be appropriately increased to improve the denoising effect. In the operator calculation module, the mapping table of the blur degree parameter β and the linear mapping parameter γ can be obtained in advance through experimental data or model training and stored in the system for query. The edge detection module can adopt various edge detection algorithms, such as the Sobel operator, the Canny operator, etc., and select the appropriate edge detection algorithm according to the actual requirements. In the edge enhancement sub-module, the fusion method of the edge information and the original infrared image can adopt various methods such as weighted fusion and superposition fusion, and the optimal fusion method can be selected according to the actual effect. Through the above specific embodiments, the parameters and configurations of the edge enhancement module 104 can be flexibly adjusted according to different application scenarios and requirements to obtain the best edge enhancement effect and system performance.

[0096] Further, the operator calculation module includes:

[0097] Mapping establishment module: configured to pre - establish a mapping table between the fuzziness parameter β and the linear mapping parameter γ, and query the mapping table according to the fuzziness parameter β to obtain the linear mapping parameter γ;

[0098] Operator calculation sub - module: configured to preset the minimum parameter L_min and the maximum parameter L_max of the Laplace operator, and calculate the Laplace operator parameter L according to the linear mapping parameter γ through the formula L = L_min+(L_max - L_min)*γ.

[0099] Among them, the mapping table reflects the corresponding relationship between the fuzziness parameter β and the linear mapping parameter γ. For example, the mapping table can establish the mapping relationship from the fuzziness parameter β to the linear mapping parameter γ in various forms such as piece - wise functions, linear functions or non - linear functions. After the fuzziness evaluation module 102 obtains the fuzziness parameter β, the mapping establishment module retrieves the linear mapping parameter γ corresponding to the fuzziness parameter β by querying the pre - established mapping table. The operator calculation sub - module is configured to receive the linear mapping parameter γ output by the mapping establishment module, and combine the preset minimum parameter L_min and maximum parameter L_max of the Laplace operator, and use the linear interpolation formula L = L_min+(L_max - L_min)*γ to calculate the Laplace operator parameter L. Among them, the minimum parameter L_min and the maximum parameter L_max of the Laplace operator are fixed values preset according to the actual application scenario and experience, representing the value range of the Laplace operator parameter L. The numerical range of the linear mapping parameter γ is usually limited between 0 and 1, and is used to control the linear interpolation of the Laplace operator parameter L within the preset range from L_min to L_max.

[0100] In some specific embodiments, the mapping table pre-established in the mapping establishment module can be configured as a linear mapping relationship. For example, assume that the value range of the fuzziness parameter β is from 0 to 100, and the value range of the linear mapping parameter γ is from 0 to 1. The mapping table can be set such that when the fuzziness parameter β is 0, the linear mapping parameter γ is 0; when the fuzziness parameter β is 100, the linear mapping parameter γ is 1. For the fuzziness parameter β between 0 and 100, the linear mapping parameter γ is calculated by linear interpolation, for example, γ = β / 100. In the operator calculation sub-module, the minimum Laplacian operator parameter L_min can be preset to 0.1, and the maximum Laplacian operator parameter L_max can be preset to 0.5. When the mapping establishment module queries the mapping table and obtains that the linear mapping parameter γ is 0.6, the operator calculation sub-module calculates the Laplacian operator parameter L as 0.34 through the formula L = 0.1 + (0.5 - 0.1) * 0.6 = 0.34. The edge detection module will perform edge detection using the Laplacian operator with the parameter of 0.34. In this way, the Laplacian operator parameter L can be adaptively adjusted according to the fuzziness of the infrared image, enabling the edge detection module to better adapt to infrared images with different fuzziness levels and improving the edge detection effect.

[0101] Further, the image fusion module 105 includes:

[0102] Normalization processing module: used to perform normalization processing on the edge-enhanced infrared image and the visible light image to obtain a normalized infrared image and a normalized visible light image;

[0103] Information entropy calculation module: used to perform image block division on the normalized infrared image and the normalized visible light image respectively to obtain normalized infrared image blocks and normalized visible light image blocks, and calculate the information entropy of each image block;

[0104] Fusion image sub-module: used to fuse the normalized infrared image blocks and the normalized visible light image blocks in a weighted fusion manner according to the information entropy of each image block to obtain a fusion image.

[0105] Among them, the normalization processing module receives the edge-enhanced infrared image and the visible light image, the output end of the module is connected to the input end of the information entropy calculation module, the output end of the information entropy calculation module is connected to the input end of the fusion image sub-module, and the fusion image sub-module outputs the fusion image.

[0106] Specifically, the normalization processing module can use a linear transformation method to process the received edge-enhanced infrared image and visible light image, and uniformly adjust the pixel value range of the two images to, for example, between 0 and 1, thereby laying a foundation for subsequent image fusion.

[0107] The information entropy calculation module can be configured to first divide the normalized infrared image and the normalized visible light image into multiple non-overlapping image blocks. For example, image blocks with a size of 16x16 pixels can be used. Then, for each image block, the information entropy value is calculated. The calculation formula of the information entropy can adopt the Shannon entropy formula, and the formula is H = -∑ p(i) *log2(p(i)), where p(i) is the probability that the pixel value i appears in the image block. By calculating the information entropy, the information richness degree of the image block is quantified.

[0108] The fused image sub-module can be configured to perform weighted fusion on the normalized infrared image block and the normalized visible light image block according to the information entropy values of each image block provided by the information entropy calculation module. As a preferred implementation manner, the information entropy value can be used as the weight coefficient, and the image block with a higher information entropy is given a higher weight in the fusion process. For example, for each corresponding infrared image block and visible light image block, the fusion weight can be set to be proportional to its information entropy value. Through weighted fusion, the image regions with rich information are more retained in the fusion result, thereby improving the quality of the fused image. Finally, all the fused image blocks are combined into a complete fused image.

[0109] In some specific implementation manners, the normalization processing module adopts the maximum-minimum normalization method to linearly map the pixel values to the range of 0 to 1. The information entropy calculation module divides the image into 16x16 pixel image blocks and calculates the information entropy of each image block using the Shannon entropy formula. The fused image sub-module adopts the weighted average fusion method, and the weight coefficient is proportional to the information entropy value of the image block, and the weight coefficient is normalized to ensure that the sum of the weight coefficients is 1. For example, the weight coefficient of the infrared image block is set to be its information entropy value divided by the sum of the information entropy values of the infrared and visible light image blocks. Through the above specific implementation manners, the weighted fusion of infrared and visible light images based on information entropy can be realized, and a high-quality fused image can be obtained.

[0110] Furthermore, the fused image sub-module includes:

[0111] The weight mapping table establishment module: used to pre-statistically calculate the information entropy range of each image block in the training image set and establish a weight mapping table between the information entropy range and the weight coefficient;

[0112] The initial weight acquisition module: used to query and obtain the initial weight coefficients of the normalized infrared image block and the normalized visible light image block according to the weight mapping table;

[0113] The image block gradient value calculation module: used to calculate the image block gradient values of the normalized infrared image block and the normalized visible light image block;

[0114] Weight coefficient adjustment module: used to adjust the initial weight coefficients of each image block according to the image block gradient value, and obtain the adjusted weight coefficients of the infrared image block and the visible light image block respectively;

[0115] The first fused image sub-module: used to fuse the normalized infrared image block and the normalized visible light image block according to the adjusted weight coefficients of the infrared image block and the visible light image block, obtain the fused image block, and then combine all the fused image blocks into a complete fused image.

[0116] Among them, the weight mapping table establishment module establishes a weight mapping table of the information entropy range and the weight coefficient in advance, and statistically calculates the information entropy range of each image block in the training image set. Thus, the corresponding relationship between the information entropy range and the weight coefficient can be determined.

[0117] The initial weight acquisition module queries the normalized infrared image block and the normalized visible light image block through the weight mapping table to obtain the initial weight coefficient, so as to provide an initial value for subsequent weight adjustment.

[0118] The image block gradient value calculation module extracts the edge and texture information of the image by calculating the image block gradient values of the normalized infrared image block and the normalized visible light image block.

[0119] The weight coefficient adjustment module uses the image block gradient value as the adjustment basis to adjust the initial weight coefficients of each image block, so that the image block with a high gradient value obtains a more accurate weight coefficient, and the edge and texture details of the image are more retained during the fusion process.

[0120] The first fused image sub-module applies the adjusted weight coefficients of the infrared image block and the visible light image block to fuse the normalized infrared image block and the normalized visible light image block, generate the fused image block, and then combine all the fused image blocks to construct a complete fused image. By adjusting the initial weight through the gradient value, the utilization rate of the local features of the image in image fusion is improved, and the image fusion effect is optimized.

[0121] Furthermore, the weight coefficient adjustment module includes:

[0122] Weight calculation module: used to calculate the adjusted weight coefficients W_IR_adjusted and W_VIS_adjusted of the infrared image block and the visible light image block through the formulas W_IR_adjusted = W_IR_initial + k * G_IR and W_VIS_adjusted = W_VIS_initial + k * G_VIS; where W_IR_initial and W_VIS_initial are the initial weight coefficients of the infrared and visible light image blocks respectively, G_IR and G_VIS are the gradient values of the infrared and visible light image blocks respectively, and k is a preset gradient weight coefficient adjustment factor.

[0123] Among them, the working principle of the weight calculation module is: based on the initial weight coefficient, adjust according to the gradient values of the infrared and visible light image blocks. If the gradient value of the image block is relatively high, it means that the image block contains rich edge or texture information, then increase the weight coefficient of the image block. In this way, the weight coefficient adjustment module can make the fusion process consider more local feature information of the image block, especially beneficial to highlighting the edge information in the image. Compared with only using the initial weight coefficient for fusion, the technical solution can adaptively adjust the weight coefficient according to the gradient information of the image block, thereby optimizing the image fusion effect, improving the quality of the fused image, and further improving the accuracy of target recognition.

[0124] Specifically, the initial weight coefficients W_IR_initial and W_VIS_initial are obtained through a pre-established weight mapping table. The weight mapping table establishes the corresponding relationship between the information entropy range and the initial weight coefficient. The gradient values G_IR and G_VIS represent the gradient values of the infrared image block and the visible light image block respectively, and the gradient values are obtained through a gradient calculation method, such as the Sobel operator or the Prewitt operator. The gradient weight coefficient adjustment factor k is a preset constant used to control the influence degree of the gradient value on the weight adjustment. For example, k can be set to 0.1.

[0125] The adjusted weight coefficients will be used for subsequent weighted fusion of image blocks. Thus, gradient information is introduced into the calculation of the weight coefficient, enabling the image fusion process to pay more attention to the edge and texture details of the image, so as to optimize the image fusion effect and quality, and provide guarantee for the subsequent improvement of target recognition accuracy.

[0126] Furthermore, the target recognition module 106 includes:

[0127] Target information acquisition module: used to process the fused image with a lightweight YOLOv5 target detection network to obtain the initial position information and category information of the human silhouette target;

[0128] Prediction module: It is used to track the human silhouette target by using the moving target tracking algorithm based on the Kalman filter according to the initial position information and category information, predict the position information of the human silhouette target in the next frame of image, and smooth the predicted position information to obtain accurate target position information;

[0129] Feedback module: It is used to generate a three-dimensional ballistic trajectory according to the accurate target position information and category information, and add the three-dimensional ballistic trajectory to the shooting feedback to feedback to the user.

[0130] Among them, through the target recognition module 106, the detection, tracking, prediction and feedback of the human silhouette target in the fused image are realized. The target recognition module 106 includes:

[0131] Target information acquisition module, which specifically uses a lightweight YOLOv5 target detection network. This network is configured to quickly process the fused image, thereby initially determining the position and category of the human silhouette target in the image.

[0132] Prediction module, connected to the target information acquisition module, receiving the initial position information and category information provided by the target information acquisition module. The prediction module uses a Kalman filter to track the motion trajectory of the human silhouette target. The Kalman filter can predict the position of the target in the next frame of image based on the target state of the previous frame and the observation result of the current frame. The prediction module also includes a smoothing function to reduce the noise in the predicted position information and improve the accuracy of position prediction.

[0133] Feedback module, receiving the accurate target position information and category information from the prediction module, and generating a three-dimensional ballistic trajectory based on this. This three-dimensional ballistic trajectory is then integrated into the shooting feedback and provided to the user.

[0134] In some specific embodiments, the target information acquisition module uses a pre-trained lightweight YOLOv5 model, which has been trained on an image dataset containing human silhouette targets and can quickly and accurately detect the target. For the Kalman filter in the prediction module, its state transition model and observation model can be adjusted according to the motion characteristics of the human silhouette target. For example, assuming the target moves at a constant speed or with a constant acceleration to improve the prediction accuracy. When the feedback module generates a three-dimensional ballistic trajectory, factors such as gravity and air resistance can be considered to make the ballistic trajectory more in line with the real physical laws. The feedback form can be various. For example, the ballistic trajectory can be displayed in a virtual scene, or the shooting result can be informed to the user through sound, force feedback, etc.

[0135] Furthermore, the feedback module includes:

[0136] Ballistic trajectory generation module: It is used to calculate the three-dimensional space vector from the shooting device to the target according to the accurate target position information and category information, combined with the virtual-real interaction system, and simulate the bullet flight trajectory according to the ballistic model to generate a three-dimensional ballistic trajectory.

[0137] Shooting effect generation module: It is used to judge whether the target is hit according to the generated three-dimensional ballistic trajectory; if it is determined that the target is hit, record the hitting position information and shooting force, and perform damage modeling on the three-dimensional model of the target according to the hitting position information and shooting force to generate a damaged effect.

[0138] Feedback sub-module: It is used to add the damaged effect to the shooting feedback and feedback it to the user.

[0139] Among them, the shooting effect generation module judges whether the target is hit based on the ballistic trajectory, records the hitting position and shooting force when hitting, and then performs damage modeling on the three-dimensional model of the target according to the hitting position and shooting force to generate a damaged effect, thereby simulating the real shooting effect.

[0140] The feedback sub-module adds the damaged effect to the shooting feedback, and finally provides the user with richer shooting feedback information including the ballistic trajectory and the damaged effect. Specifically, the ballistic trajectory generation module can adopt a physical ballistic model, such as a parabola model considering factors such as air resistance and gravity, to more accurately simulate the bullet flight trajectory. In the shooting effect generation module, the damage modeling process can generate different degrees of damaged effects in different regions of the three-dimensional model of the target according to different hitting positions, such as generating different damaged models when hitting the head and torso. The feedback sub-module can superimpose the ballistic trajectory and the damaged effect on the virtual-real interaction scene in a visual way, such as displaying the ballistic trajectory and the target damage process in the form of animation or special effects.

[0141] In some specific embodiments, the ballistic trajectory generation module adopts a simplified linear ballistic model to ensure the real-time generation of the ballistic trajectory when the resources of the edge computing device are limited. The shooting effect generation module pre-stores various damaged models of the target with different degrees and positions, and quickly selects and loads the corresponding damaged model according to the hitting position and shooting force, thereby reducing the calculation amount of damage modeling and improving the real-time performance of the feedback. The feedback sub-module superimposes the three-dimensional ballistic trajectory on the virtual scene with a red dotted line, and updates the damaged effect of the target to the three-dimensional model of the target in the form of a texture in real time, enabling the user to intuitively observe the shooting effect. Thus, the enhancement of the shooting feedback effect is realized under the condition of limited edge computing resources.

[0142] In this document, relational terms such as first and second are used solely to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations.

[0143] The above description is only for the embodiments of the present application and is not intended to limit the scope of protection of the present application. For those skilled in the art, the present application may have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the scope of protection of the present application.

Claims

1. An AI vision-based dynamic image recognition system for virtual-real interactive shooting training of light-absorbing targets in indoor low-light environments, characterized in that, The system includes: An image acquisition module: used to obtain a visible light image and an infrared image containing a human silhouette target synchronously acquired by a visible light camera and an infrared camera; A blur degree evaluation module: used to evaluate the blur degree of the infrared image and obtain a blur degree parameter β; An enhancement coefficient calculation module: used to calculate an adaptive edge enhancement coefficient α according to the blur degree parameter β; An edge enhancement module: used to perform edge detection on the infrared image using a Laplacian operator to obtain infrared image edge information, and enhance the infrared image edge information according to the adaptive edge enhancement coefficient α, and obtain an edge-enhanced infrared image; An image fusion module: used to perform pixel-level fusion on the edge-enhanced infrared image and the visible light image to obtain a fused image; A target recognition module: used to perform target detection and recognition on the fused image, obtain the target position and category information of the human silhouette target, and provide shooting feedback to the user according to the target position and the category information; The blur degree evaluation module includes: A gradient calculation module: used to calculate the image gradient values of multiple frames of the infrared image; An image sharpness evaluation module: used to calculate an image sharpness evaluation index according to multiple image gradient values: σ = ∑(G - μ)^2 / N, where σ is the image sharpness evaluation index, G is the image gradient value, μ is the average image gradient, and N is the total number of image pixels; A blur parameter calculation module: used to calculate the mean value of the image sharpness evaluation index σ of multiple frames of infrared images and use this mean value as the blur degree parameter β; The enhancement coefficient calculation module includes: A preset coefficient acquisition module: used to acquire a preset minimum enhancement coefficient α_min, a maximum enhancement coefficient α_max, and a maximum blur degree index threshold β_max; An enhancement coefficient determination module: used to use the maximum enhancement coefficient α_max as the adaptive edge enhancement coefficient α when the difference between the blur degree parameter β and β_max is less than a first preset difference; When the difference between the blur degree parameter β and 0 is less than the first preset difference, use the minimum enhancement coefficient α_min as the adaptive edge enhancement coefficient α; When the differences between the blur degree parameter β and 0 and between β and β_max are both greater than or equal to the first preset difference, the adaptive edge enhancement coefficient α is calculated by the formula α = α_min + (α_max - α_min) * (β / β_max).

2. The AI vision-based dynamic image recognition system according to claim 1, wherein, The edge enhancement module includes: A noise reduction processing module: used to perform Gaussian filtering on the infrared image, perform convolution on the Gaussian kernel and the infrared image to obtain a noise-reduced infrared image I_IR_filtered; An operator calculation module: used to calculate an adaptive Laplacian operator according to the blur degree parameter β; An edge detection module: used to perform edge detection on the noise-reduced infrared image I_IR_filtered using the Laplacian operator to obtain the infrared image edge information; Edge Enhancement Sub-module: It is used to fuse the edge information of the infrared image and the original infrared image according to the adaptive edge enhancement coefficient α to obtain the edge-enhanced infrared image I_IR_enhanced.

3. An AI vision-based dynamic image recognition system according to claim 2, characterized in that, The operator calculation module includes: Mapping Establishment Module: It is used to pre-establish a mapping table between the blur degree parameter β and the linear mapping parameter γ, and query the mapping table according to the blur degree parameter β to obtain the linear mapping parameter γ; Operator Calculation Sub-module: It is used to preset the minimum parameter L_min and the maximum parameter L_max of the Laplace operator, and calculate the Laplace operator parameter L through the formula L = L_min+(L_max - L_min)*γ according to the linear mapping parameter γ.

4. An AI vision-based dynamic image recognition system according to claim 1, characterized in that, The image fusion module includes: Normalization Processing Module: It is used to perform normalization processing on the edge-enhanced infrared image and the visible light image to obtain a normalized infrared image and a normalized visible light image; Information Entropy Calculation Module: It is used to perform image block division on the normalized infrared image and the normalized visible light image respectively to obtain a normalized infrared image block and a normalized visible light image block, and calculate the information entropy of each image block; Fusion Image Sub-module: It is used to fuse the normalized infrared image block and the normalized visible light image block in a weighted fusion manner according to the information entropy of each image block to obtain the fusion image.

5. The AI vision-based dynamic image recognition system according to claim 4, wherein The fusion image sub-module includes: Weight Mapping Table Establishment Module: It is used to pre-statistically calculate the information entropy range of each image block in the training image set and establish a weight mapping table between the information entropy range and the weight coefficient; Initial Weight Obtaining Module: It is used to query and obtain the initial weight coefficients of the normalized infrared image block and the normalized visible light image block according to the weight mapping table; Image Block Gradient Value Calculation Module: It is used to calculate the image block gradient values of the normalized infrared image block and the normalized visible light image block; Weight Coefficient Adjustment Module: It is used to adjust the initial weight coefficients of each image block according to the image block gradient value to obtain the adjusted weight coefficients of the infrared image block and the visible light image block respectively; First Fusion Image Sub-module: It is used to fuse the normalized infrared image block and the normalized visible light image block according to the adjusted weight coefficients of the infrared image block and the visible light image block to obtain the fusion image block, and then combine all the fusion image blocks into the complete fusion image.

6. The AI vision-based dynamic image recognition system according to claim 5, characterized in that, The weight coefficient adjustment module includes: Weight calculation module: used to calculate the adjusted weight coefficients W_IR_adjusted of infrared image blocks and W_VIS_adjusted of visible light image blocks through the formulas W_IR_adjusted = W_IR_initial + k * G_IR and W_VIS_adjusted = W_VIS_initial + k * G_VIS; where W_IR_initial and W_VIS_initial are the initial weight coefficients of infrared and visible light image blocks respectively, G_IR and G_VIS are the gradient values of infrared and visible light image blocks respectively, and k is a preset gradient weight coefficient adjustment factor.

7. An AI vision-based dynamic image recognition system according to claim 1, characterized in that The target recognition module includes: Target information acquisition module: used to process the fused image with a lightweight YOLOv5 target detection network to obtain the initial position information and category information of the human silhouette target; Prediction module: used to track the human silhouette target using a moving target tracking algorithm based on a Kalman filter according to the initial position information and the category information, predict the position information of the human silhouette target in the next frame of image, and smooth the predicted position information to obtain accurate target position information; Feedback module: used to generate a three-dimensional ballistic trajectory according to the accurate target position information and the category information, and add the three-dimensional ballistic trajectory to the shooting feedback to feedback to the user.

8. An AI vision-based dynamic image recognition system according to claim 7, wherein, The feedback module includes: Ballistic trajectory generation module: used to calculate the three-dimensional space vector from the shooting device to the target according to the accurate target position information and the category information, combined with the virtual-real interaction system, simulate the bullet flight trajectory according to the ballistic model, and generate a three-dimensional ballistic trajectory; Shooting effect generation module: used to judge whether the target is hit according to the generated three-dimensional ballistic trajectory; if it is determined that the target is hit, record the hit position information and the shooting force, and perform damage modeling on the three-dimensional model of the target according to the hit position information and the shooting force to generate a damaged effect; Feedback sub-module: used to add the damaged effect to the shooting feedback to feedback to the user.

Images