Intelligent edge calculation method for fire target detection

Abstract

The invention discloses an intelligent edge calculation method for fire target detection, and belongs to the technical field of fire target detection, and the method comprises the following steps: S1, making a fire target detection data set; s2, an improved YOLO-AFPN-LADH-WIOU model is formed, and an improved YOLO-AFPN-LADH-WIOU S3, carrying out evaluation and transverse comparison on the improved YOLO-AFPN-LADH-WIOU model, and carrying out visual analysis on a result; s4, based on the step S3, realizing remote deployment of the basic model on the development board; s5, establishing a connection between the edge detection system and the client monitoring end by using a TCP transmission control protocol so as to perform interaction; and S6, realizing image acquisition and user GUI display. According to the intelligent edge calculation method for fire target detection, the precision, the speed and the breadth of detection are improved, and meanwhile, the real-time performance and the accuracy of fire monitoring and protection in practice are ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of fire target detection, and in particular to an intelligent edge computing method for fire target detection. Background Art

[0002] Forest fires are a serious environmental problem. They not only endanger people's lives and property safety, but also emit a large amount of carbon-containing gases and produce huge carbon emissions. In order to meet the challenges of forest fires, many countries and organizations have taken measures such as prevention, monitoring and fire fighting. This includes enhancing forest management and monitoring capabilities, improving the accuracy and response speed of fire warning systems, and strengthening research on fire monitoring and fire fighting technologies.

[0003] In the monitoring industry, traditional fire prevention detection mainly relies on contact smoke sensors, which use the principle that the resistance of smoke-sensitive elements changes due to the influence of smoke concentration to send analog signals corresponding to the smoke concentration to the host to achieve fire warning. Since the detection equipment of contact sensors has a slow response speed, it must be installed near the fire point and can only be used in smaller spaces and in very limited environments. Contact smoke sensors can usually only detect areas where smoke directly contacts the sensor. There may be blind spots in hidden areas or situations where smoke diffuses slowly, resulting in delayed fire alarms. In addition, traditional smoke sensors are easily interfered by other factors such as dust and steam, which may lead to false fire alarms; at the same time, in certain fire scenarios (such as rapid combustion, high temperature, etc.), the sensor may miss reports, affecting the timely detection and handling of fires.

[0004] Compared with contact smoke sensors, image-based flame and smoke detection technology can cover a wider area and detect hidden areas and places where smoke spreads slowly. The forest environment monitoring method based on deep learning avoids the complex image preprocessing process and can automatically learn the high-level feature representation of disasters, greatly improving the accuracy of detection. Image-based detection technology can achieve remote monitoring through the network, allowing users to monitor fire and smoke conditions anytime and anywhere, and achieve more convenient management and control. Therefore, there is great potential in designing a remote fire monitoring system based on deep learning. Summary of the invention

[0005] The purpose of the present invention is to provide an intelligent edge computing method for fire target detection, which improves the accuracy, speed and breadth of detection, while ensuring the real-time and accuracy of fire monitoring and protection in practice.

[0006] To achieve the above object, the present invention provides a fire target detection intelligent edge computing method, comprising the following steps: S1. Create a fire target detection dataset; S2. Select the YOLOv8 target detection model as the basic architecture, integrate the improvement strategy, form the improved YOLO-AFPN-LADH-WIOU model, and train the improved YOLO-AFPN-LADH-WIOU model; S3. Evaluate and compare the above improved YOLO-AFPN-LADH-WIOU model, and visualize the results; S4. Based on step S3, realize remote deployment of the improved YOLO-AFPN-LADH-WIOU model on the development board; S5. Use TCP transmission control protocol to establish a connection between the edge detection system and the client monitoring terminal for interaction; S6. Realize image acquisition and user GUI display.

[0007] Preferably, the data set in step S1 includes positive samples and negative samples. The positive samples are specifically selected and produced with a variety of fire pictures, including flame and smoke pictures of different sizes in various environments, and the pictures are subjected to data enhancement methods such as adding noise and flipping; the negative samples include industrial smoke, sunset glow, clouds, haze, morning fog, lights, sun, red flowers and background images with similar flame or smoke features, and then the positive and negative samples are trained together.

[0008] Preferably, the improvement strategy in step S2 is specifically the AFPN target detection network structure, the LADH detection head and the WIoU loss function.

[0009] The impact of each improvement strategy is: AFPN effectively integrates the features of the bottom, top and bottom layers through a progressive feature fusion strategy to reduce the semantic gap between features at different levels and improve the feature fusion effect, thereby improving the target detection performance. Integrating AFPN into the YOLOv8 neck effectively reduces the information gap between non-adjacent layers and promotes improved feature fusion. The LADH detection head improves the original detection head to realize a fast and lightweight target detection module, which is suitable for complex background and small target detection tasks in remote sensing images. The LADH detection head is used in the model to improve the detection accuracy while ensuring the rapid response capability of the model, meeting the needs of real-time processing. The WIoUv3 function is used to improve the performance of the target detection model when dealing with challenging scenes. Using the WIoUv3 loss function, the model's detection ability for small targets and difficult-to-detect targets is improved.

[0010] Preferably, in step S2, the improved YOLO-AFPN-LADH-WIOU model is trained as follows: Integrate the asymptotic feature pyramid network (AFPN) structure to improve target detection performance; The lightweight asymmetric detection head (LADH) loss function is used to reduce the mutual interference between tasks; The WIoU loss function is used to improve the accuracy of model detection.

[0011] Preferably, in step S3, the model is evaluated and analyzed by multiple indicators such as accuracy, recall, average accuracy mean, and model weight scale, and a comprehensive analysis is performed using confusion matrix, model inference results, and heat map.

[0012] Preferably, step S4 connects the PC and the development board via IP, and uses the VNC viewer service to realize the remote display of the development board screen; installs the OPENCV library on the development board, and transmits the basic model weight file and other configuration files from the PC to the development board, thereby realizing the inspection of the fire target by using the development board as an edge computing system.

[0013] Preferably, step S5 creates a communication channel by creating a Socket keyword in the network to connect the server and the client, and uses the TCP protocol to conduct a dialogue between the two ends.

[0014] Preferably, the image acquisition part in step S6 includes an external camera, which is connected to the edge detection system. The captured picture is directly identified and processed on the edge detection system, and the processed result is then transmitted to the customer monitoring segment and displayed on the client through a self-made GUI interface. The GUI interface includes a detected picture frame, a detection target type result frame, a status reminder frame, and a detection record frame.

[0015] Therefore, the present invention adopts the above-mentioned fire target detection intelligent edge computing method, and the beneficial effects are as follows: The present invention uses a fire dataset with flames and smoke of different sizes in various environments for training. The dataset includes fire pictures that have been noise-processed and contain occluders. This makes the model more adaptable to forest scenes with multiple interferences, high noise, and small target fires, and can improve the generalization ability of the model, enabling it to accurately identify fires in different environments.

[0016] The present invention adopts the improved YOLO-AFPN-LADH-WIOU model, which can achieve more efficient and accurate recognition compared with the traditional target detection algorithm. The improved model has higher speed and accuracy in target detection tasks. By using this model, a large number of images can be detected in real time in a short time, and signs of fire can be quickly found, which improves the efficiency and accuracy of detection.

[0017] Based on the idea of ​​embedded development, the present invention realizes the remote deployment of the model to the monitoring end system, so as to directly identify and process the collected image data, and then transmit and display the processed images and results to the client monitoring end in real time. Through such a process, the time of first transmitting and then processing the image at the back end can be effectively reduced, ensuring the real-time and urgency requirements of fire monitoring, which is conducive to the monitoring personnel to observe and take treatment measures in time.

[0018] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a front view of an embodiment of an intelligent edge computing method for fire target detection of the present invention; Figure 2 It is a flow chart of intelligent fire target detection according to an embodiment of an intelligent edge computing method for fire target detection of the present invention; Figure 3 This is a diagram of the training results of an intelligent fire monitoring model according to an embodiment of an intelligent edge computing method for fire target detection of the present invention; Figure 4 It is a confusion matrix diagram of an embodiment of a fire target detection intelligent edge computing method of the present invention; Figure 5 This is an example diagram of detection results of an embodiment of an intelligent edge computing method for fire target detection of the present invention; Figure 6 This is a user-side monitoring screen diagram of an embodiment of an intelligent edge computing method for fire target detection in the present invention. DETAILED DESCRIPTION

[0020] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0021] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0022] Embodiment 1 like Figure 1 As shown, the present invention provides an intelligent edge computing method for fire target detection. Figure 2This is the flow chart of intelligent fire target detection of the present invention. The fire image data from various places in China and abroad were used for experiments, and a total of 12,549 fire images and 1,034 non-fire images were obtained. According to different categories, 80% of the 13,583 images were selected as training samples and 20% as test samples. The model pre-training weight uses YOLOv8n.pt, the batch-size is set to 16, 200 epochs are set for each experiment, and the initial learning rate is 0.01. The specific steps are as follows: Step 1: Create a fire target detection training dataset to enhance the capacity of the dataset and the robustness and generalization ability of the model, so that the model can be applied to fire target detection in high-noise and multi-interference environments. This dataset contains images of fires in many other different scenarios, such as factories, indoors, and schools, which improves the robustness of the model and effectively enhances the adaptability of the model in practical applications.

[0023] Since the basic model has poor detection effect on small targets, it is prone to misjudgment and missed detection, and the system is often deployed at high places, so the targets are often small. Considering the complex and changeable natural environment in reality, the fire points of early forest fires are small and weak in intensity, and they are often unclear small targets. In order to improve the accuracy of the model's detection of small targets, some original images are enhanced, such as adding noise, flipping and other series of operations.

[0024] Secondly, in order to enable the model to accurately detect fires at different distances, three types of flame or smoke images, large, middle, and small, were selected based on the size of the flame and smoke in the image, thereby enhancing the accuracy of the model in detecting fires at different distances. This classification not only helps improve the model's ability to detect fires of different scales, but also enhances the model's performance in a variety of scenarios.

[0025] In addition, we selected some background images with similar features to flames or smoke, including industrial chimneys, sunset glow, clouds, haze, morning fog, lights, the sun, red flowers, and red birds. We set the “other” label on some of them for distinction, and the rest are used as negative sample data for this dataset.

[0026] Step 2: Improve the basic model by integrating the AFPN target detection network structure, LADH detection head, and WIoUv3 loss function.

[0027] The convolutional neural network model designed in the present invention is improved on the basis of the entry framework YOLOv8 for target detection. In terms of key innovation and enhancement in the model architecture, three improvement strategies, namely the Asymptotic Feature Pyramid Network (AFPN), the Lightweight Asymmetric Detection Head (LADH) and the WIoU loss function, are synthesized and integrated to form an improved YOLO-AFPN-LADH-WIOU model.

[0028] Step 2.1: Integrate the Asymptotic Feature Pyramid Network (AFPN) to improve object detection performance.

[0029] AFPN is a new network structure for object detection. It effectively integrates the features of the bottom, top and top layers through a progressive feature fusion strategy to reduce the semantic gap between features at different levels, improve the feature fusion effect, and thus improve the object detection performance. The core idea of ​​AFPN is to reduce the semantic gap between different levels by gradually integrating the features of higher layers starting from the adjacent low-level features through a progressive feature fusion strategy. AFPN also introduces an adaptive spatial fusion mechanism, which enhances the importance of key levels and suppresses the contradictory information between different objects through changing spatial weights in multi-level feature fusion. In addition, AFPN uses the alignment of the bottom-level features to ensure that the semantics of features at different levels gradually approach each other during the fusion process, thereby improving the overall feature fusion effect. This fusion strategy not only enhances the representation ability of the features, but also improves the performance of object detection, especially when dealing with small and multi-scale objects. The YOLO-AFPN-LADH-WIOU model integrates AFPN into the YOLOv8 neck, which effectively reduces the information gap between non-adjacent layers and promotes improved feature fusion.

[0030] Step 2.2: Use Lightweight Asymmetric Detection Head (LADH) to reduce mutual interference between tasks.

[0031] The core idea of ​​the LADH detection head is to decouple the classification and regression subtasks in the detection task and process them using different network paths to reduce the mutual interference between them. Compared with the standard convolution used in the original YOLO algorithm, the LADH detection head uses 3×3 depthwise separable convolution (DWConv). This is a convolution method with fewer parameters and computational complexity. It effectively reduces the number of model parameters by decomposing the convolution operation into two steps: depthwise convolution and pointwise convolution, while maintaining the ability to extract features.

[0032] Secondly, the LADH detection head uses three different channels to handle classification, regression, and IoU (intersection over union) prediction tasks respectively. In the depthwise separable convolution, the depthwise convolution performs convolution operations on each input channel independently, while the pointwise convolution is used to mix the information of different channels to generate a new feature map. This method not only reduces the computational complexity of the model, but also improves the model's ability to detect small targets, because small targets may only occupy a few pixels in remote sensing images and require more refined feature extraction.

[0033] In addition, the design of the LADH detection head also takes into account the lightweight of the model, which is particularly important for deploying models on devices with limited computing resources. By reducing model parameters and computational complexity, the LADH detection head helps achieve faster inference speed while maintaining high detection accuracy. This design is not only suitable for specific application scenarios of remote sensing images, but also provides valuable reference for other fields that require lightweight models.

[0034] Step 2.3: Use WIoU loss function to improve the accuracy of model detection.

[0035] WIoUv3 (Wise-IoU version 3) is an advanced bounding box regression loss function based on the traditional IoU (intersection over union) loss, but by introducing several key innovations to enhance the model's learning ability. WIoUv3 first assigns weights to each category and uses different weights to measure the overlap between different categories, thereby reducing the difference between categories and their impact on test results. This category weighting mechanism enables the model to focus more on the features of specific categories, improving the accuracy of detection.

[0036] WIoUv3 introduces a dynamic non-monotonic focusing coefficient' r ', this coefficient is based on the abnormality parameter βThis focusing mechanism can dynamically evaluate the quality of the anchor frame and reduce the impact of bad gradients by gradient amplification to ensure the high quality of the anchor frame. In this way, the model can focus more on optimizing those high-quality anchor frames during training, while giving smaller gradient gains to low-quality anchor frames to avoid their negative impact on model training. The formula is as follows: ; ; ; ; ; in, express Loss function; express Loss function; represents the penalty term; Represents the predicted box loss; and Respectively represent the horizontal and vertical coordinates of the center point of the real frame; and Respectively represent the horizontal and vertical coordinates of the center point of the prediction box; and Respectively represent the width and height of the smallest rectangle containing the predicted box and the true box; the superscript " " represents the computational graph and separation; β It is defined as the outlier degree, which is used to indicate a Compared to most the degree of deviation; is the non-monotonic focusing coefficient; α and δ is a hyperparameter.

[0037] Smaller outliers mean higher quality anchor boxes, and a small gradient gain is assigned to focus bounding box regression on the normal quality of the anchor boxes. Anchor boxes with larger outliers are assigned a smaller gradient gain to effectively prevent low-quality data from generating significant harmful gradients. β = δ hour, r = 1. If an anchor's outlier box satisfies β=C (where C is a constant), then the anchor box will receive the highest gradient. Since the quality criteria of LIOU (LogarithmicIntersection over Union) and anchor boxes are dynamically changing, WIOUv3 is able to dynamically assign the gradient gain strategy that best suits the current situation.

[0038] Step 3: Select appropriate parameter indicators to evaluate and compare the models In order to verify the detection performance of the model, multiple indicators such as accuracy, recall, mean average precision and model weight scale were used to evaluate and analyze the model.

[0039] Accuracy (P) measures the accuracy of the model, which is how many of the positive categories predicted by the model are true positive categories: ; Among them, TP is the number of positive samples classified correctly, and FP is the number of positive samples classified incorrectly; If the threshold of the overlap (IOU) between the predicted box and the true box is 0.7, when the IOU value of a predicted box and a true box is greater than the threshold, the prediction result is judged as a true positive (TP), otherwise it is judged as a false positive (FP): Recall (R) measures the comprehensiveness of the model, which is how many of all true positive classes are successfully detected by the model: ; Among them, FN is the number of negative samples that are misclassified; The average accuracy (AP) can measure the detection performance of the model and is defined as the detection accuracy of each sample.

[0040] ; Among them, the higher the AP value, the better the detection performance of the model; The mean average accuracy (mAP) is a comprehensive evaluation indicator that comprehensively considers the accuracy of the model in different categories and is the mean of AP of all categories: ; Among them, mAP (mean AP) is the average of various AP types, and is the value obtained by averaging each type of AP. The value range of mAP is in the interval [0,1], and the larger the value, the better.

[0041] Figure 3In order to improve the F1 change curve, P change curve, PR change curve and R change curve of the model, the parameters and detection performance of the model can be judged more intuitively.

[0042] In addition, in order to intuitively demonstrate the detection performance of the improved model, we use confusion matrix, model inference results and heat map for comprehensive analysis. These tools not only enhance our data visualization capabilities, but also enable us to clearly compare the detection effects of different models.

[0043] The rows and columns of the confusion matrix represent the actual categories and predicted categories, respectively, providing an intuitive way to observe the predictive ability of the model. Therefore, the confusion matrix of the improved model is generated, such as Figure 4 As shown in the figure, it can effectively visualize the performance of the two models on the fire dataset. The dataset used by the model contains three types of targets: fire, smoke, and other. In the confusion matrix, the values ​​on the diagonal represent the correct prediction ratio of each category, reflecting the accuracy of the model on each category. The values ​​in other areas of the matrix reveal the proportion of incorrectly predicted categories, reflecting the shortcomings of the model on certain specific categories, such as some categories may be misclassified as other categories.

[0044] The present invention conducted an inference experiment and selected a number of representative fire scenes. These scenes include forests, urban roads, public facilities and traffic intersections, etc. All images include flames or smoke, which are suitable for detecting fire targets. In addition to selecting some common fire scenarios, we also selected some large, small and noise-interfered fire target images. By using the improved model to detect these images, we visualized the detection of this model in different scenarios, small targets, and multiple noise-interfered fires. Example diagram of the detection results of the intelligent fire monitoring model is shown in the figure below. Figure 5 shown.

[0045] Step 4: Implement remote deployment of the improved YOLO-AFPN-LADH-WIOU model on the development board The PC and development board are connected via IP, and the VNC viewer service is used to realize the remote display of the development board screen; the OPENCV library is installed on the development board, and the model weights and configuration files are transferred from the PC to the development board.

[0046] Step 5: Establish a connection between the edge detection system and the client monitoring terminal for interaction. The client monitoring screen is shown in the figure below: Figure 6 shown.

[0047] The present invention connects the server and the client by creating a Socket keyword in the network, creates a communication channel, and uses the TCP protocol to conduct dialogue between the two ends. Socket is a mechanism that allows different computers to exchange data, and is also an abstract layer that allows applications to communicate through the network. It establishes a connection between computers so that they can exchange data. The process is as follows: First, the edge processing system creates a socket and connects to the server. This process includes binding the socket to the remote server to establish a communication channel; Next, the edge processing system sends a data request and reads the server's response data. This process continues until all data is exchanged; After completing the data transmission, the edge processing system will close the connection and end the TCP conversation. The remote receiving server will first initialize the Socket, establish a stream socket, and bind it to the specific address and port of the local machine; After completing these steps, the server notifies TCP that it is ready to receive connections and calls the accept() method to enter a blocked state, waiting for a connection request from the edge processing system. Once the edge processing system successfully establishes a connection with the server, the edge processing system sends a data request. After receiving the request, the server processes it accordingly and returns the processing result to the edge processing system as response data. The edge processing system then reads the data to ensure that all data exchanges are completed; Finally, both parties close the connection, marking the end of the interaction.

[0048] Step 6: Realize image acquisition, transmission and GUI display.

[0049] Table 1. Results of ablation experiment of intelligent fire monitoring model ; The experimental results show that when applied to the basic model, each improvement strategy will improve the detection performance to varying degrees. In addition, among all the above models, the YOLO-AFPN-LADH-WIOU model performs better in terms of accuracy (P) and model weight scale (Scale). Compared with the basic model, the model YOLOv8-AFPN after using the AFPN network has improved its accuracy by 0.7%, and the model weight scale has dropped significantly to 4.768MB, but its accuracy is still lower than the final improved model; the model YOLOv8-LADH model after introducing the LADH detection head has greatly improved its accuracy, but there is still a certain gap in the model weight size compared to the final improved model. After applying the three strategies to the YOLOv8n model at the same time, the model accuracy is improved by 1% compared with the basic model, and the model weight scale is also reduced from 6.097MB to 4.861MB. The accuracy of fire target detection is high, and the model weight is relatively low, with the best overall performance.

[0050] Obviously, the YOLO-AFPN-LADH-WIOU model is applied to achieve the best value of model performance. Considering the high accuracy and real-time requirements of fire detection, a part of mAP50 is sacrificed to achieve higher detection accuracy and lower model weight size, so as to reduce the incidence of missed detection and false detection and effectively improve detection efficiency.

[0051] Table 2 Comparison experimental results of intelligent fire monitoring models ; The results in the table show that compared with other mainstream models, the YOLO-AFPN-LADH-WIOU proposed in this invention has better detection performance. Compared with the YOLOv5s model, the accuracy and mAP0.5 of the model proposed in this invention are increased by 3.3% and 7% respectively, proving that the model more accurately identifies the correct target when detecting the target, and the target detection accuracy is higher; compared with the YOLOv8n and YOLOV8-RepNCSPFPN models, the YOLO-AFPN-LADH-WIOU model not only has the highest accuracy (92.1%) and mAP@0.5 (88.8%), but also has the lowest model weight size (4.861 MB). This result shows that although the model is more complex, its structure still maintains its lightweight characteristics; compared with the YOLOV8n-AFPN-LADH-VariFocalLoss model and the YOLOV8n-AFPN-LADH-XIoU model, the YOLO-AFPN-LADH-WIOU model has significantly improved in accuracy and mAP@0.5, showing stronger target detection capabilities.

[0052] This shows that based on the introduction of the AFPN network architecture and the LADH detection head, the use of the WIoU loss function can better optimize the accuracy and model size of the model compared to the VariFocalLoss and XIoU loss functions, ensuring that the model is lightweight while also being able to complete fire target detection with high precision. Our model not only has the highest accuracy and mAP50, but also has the lowest model weight, which is attributed to the fact that the invention improves the accuracy and speed of the model's fire target detection by introducing three improved methods: the AFPN network architecture, the LADH detection head, and the WIoU loss function, thereby improving the overall performance of the model.

[0053] In summary, from the perspective of the model, the YOLO-AFPN-LADH-WIOU model has advantages in accuracy, average accuracy and model weight scale. This shows that in the flame and smoke detection tasks, the model that introduces the AFPN network architecture and replaces it with the LADH detection head and WIoU loss function can improve the detection performance and has a lighter model structure. The improved YOLO-AFPN-LADH-WIOU model of the present invention has better performance in both flame and smoke detection.

[0054] The present invention first collects image data through a camera and obtains a frame image. By reading the frame image, a real-time target detection algorithm is used to detect fire targets: if the target is detected, an abnormality is displayed, and the flame or smoke position and its confidence are calibrated; otherwise, the detection result is that no fire is detected. The detected image data is then transmitted to the receiving end through network communication (TCP), and the receiving end displays the result. The purpose is to calibrate the information of the target in the monitoring image by inputting the image prediction bounding box and category probability, so as to detect smoke and flames in actual scenes.

[0055] Therefore, the present invention adopts the above-mentioned intelligent edge computing method for fire target detection. This method utilizes the advantages of deep learning and target detection algorithms, designs an image-based non-contact detection method, and improves the accuracy, speed and breadth of detection; based on the basic model of the target detection algorithm, a variety of improvement strategies are introduced to improve the detection accuracy of the improved model and reduce the weight scale; then combined with embedded development, the remote deployment of the model is realized, so that the edge detection end and the monitoring end can work together efficiently, ensuring the real-time and accuracy requirements of fire monitoring and protection in practice.

[0056] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A fire target detection intelligent edge computing method, characterized in that: The following steps are involved: S1. Create a fire target detection dataset; S2. Select the YOLOv8 target detection model as the basic architecture, integrate the improvement strategy, form the improved YOLO-AFPN-LADH-WIOU model, and train the improved YOLO-AFPN-LADH-WIOU model; S3. Evaluate and compare the above improved YOLO-AFPN-LADH-WIOU model, and visualize the results; S4. Based on step S3, realize remote deployment of the improved YOLO-AFPN-LADH-WIOU model on the development board; S5. Use TCP transmission control protocol to establish a connection between the edge detection system and the client monitoring terminal for interaction; S6. Realize image acquisition and user GUI display.

2. According to claim 1, a fire target detection intelligent edge computing method is characterized by: The data set in step S1 includes positive samples and negative samples. The positive samples are specifically selected and produced with a variety of fire pictures, including flames and smoke pictures of different sizes in various environments, and the pictures are subjected to data enhancement methods such as adding noise and flipping; the negative samples include industrial exhaust smoke, sunset glow, clouds, haze, morning fog, lights, sun, red flowers and background images with similar flame or smoke features, and then the positive and negative samples are trained together.

3. The fire target detection intelligent edge computing method according to claim 2 is characterized in that: The improvement strategy in step S2 is specifically the AFPN target detection network structure, LADH detection head and WIoU loss function.

4. The fire target detection intelligent edge computing method according to claim 3 is characterized in that: In step S2, the improved YOLO-AFPN-LADH-WIOU model is trained as follows: Integrate AFPN structure to improve target detection performance; Use LADH loss function to reduce mutual interference between tasks; The WIoU loss function is used to improve the accuracy of model detection.

5. The fire target detection intelligent edge computing method according to claim 4 is characterized in that: In step S3, the model is evaluated and analyzed by multiple indicators such as accuracy, recall, average accuracy and model weight scale, and a comprehensive analysis is performed using confusion matrix, model inference results and heat map.

6. The fire target detection intelligent edge computing method according to claim 5 is characterized in that: Step S4 connects the PC and the development board through IP, and uses the VNC viewer service to realize the remote display of the development board screen; installs the OPENCV library on the development board, and transmits the basic model weight file and other configuration files from the PC to the development board, so as to realize the inspection of the fire target by using the development board as an edge computing system.

7. The fire target detection intelligent edge computing method according to claim 6 is characterized by: Step S5 connects the server and the client by creating a Socket keyword in the network, creates a communication channel, and uses the TCP protocol to conduct dialogue between the two ends.

8. The fire target detection intelligent edge computing method according to claim 7 is characterized in that: The image acquisition part in step S6 includes an external camera, which is connected to the edge detection system. The captured image is directly identified and processed on the edge detection system, and the processed result is then transmitted to the customer monitoring segment and displayed on the client through a self-made GUI interface. The GUI interface includes a detected image frame, a detection target type result frame, a status reminder frame, and a detection record frame.

Images