A fire detection method based on an anchor-free network

Abstract

The present invention relates to a fire detection method based on an anchor-free network, comprising the following steps: Image acquisition: obtaining a real-time video stream through a camera and converting it into an image format in a computer; Foreground judgment; Network detection, if it is judged that there is a suspicious moving foreground target in a certain frame of image, then send this frame of image into a trained fire detection network for judging the flame target. First, feature extraction is carried out through a backbone network, then feature fusion is carried out by using an optimized fusion network, and finally, the category and position information of the flame target are predicted through two branches of classification and regression of the detection network; Alarm response.

Description

Technical Field

[0001] The present invention belongs to the field of fire detection in image processing technology, and particularly relates to a fire detection method based on an anchor-free network. Background Art

[0002] Fire is one of the common disasters with great destructiveness in daily life, posing a great threat to people's lives and property safety and social harmony and stability. Since the heat released by the flame in the initial stage of combustion is small, it moves slowly, and the affected area is small. If it is detected and measures are taken in time at the initial stage of the fire, many tragedies can be avoided. Therefore, real-time and accurate fire detection technology has very important practical application value.

[0003] Most traditional fire detection technologies use sensors as carriers to judge whether a fire occurs by detecting the parameter changes during the flame combustion process. Although the principle of this technology is simple, there are many problems in the actual application process, such as small coverage area, large influence by the environment, short service life, etc. With the construction of smart cities in China and the rapid development of digital technology, the image-based fire detection technology based on video surveillance platforms has emerged as the times require. This technology has low requirements for the environment, is applicable to various scenarios in real life, has high detection accuracy, short response time, and can also provide rich fire alarm information, providing strong support for subsequent rescue and analysis.

[0004] Early image-based fire detection technologies mainly used the method of manually extracting features and combined with the idea of machine learning to judge the extracted features. However, the method of manually extracting features has poor generalization ability and is difficult to adapt to various complex scenarios in the real environment. In recent years, the field of deep learning has been widely studied and developed. The fire detection technology based on deep learning can realize the automatic extraction of features, thereby obtaining richer feature information, significantly improving the accuracy of fire detection in the actual scenario, and becoming the key research content in the current field of intelligent fire protection. However, most of the current fire detection technologies based on deep learning have problems such as high model complexity, low detection accuracy of flame targets in complex backgrounds, and easy false alarms, which to a certain extent affect the detection effect of flame targets in actual applications, and the applicability to fire detection tasks in real scenarios needs to be further improved. Summary of the Invention

[0005] Aiming at the limitations and deficiencies of the prior art, the present invention fully considers the characteristics of flame targets and actual detection requirements, and provides a new fire detection method based on an anchor-free convolutional neural network, which can realize a real-time and accurate fire detection process and fully meet the needs of actual fire detection tasks. The technical solution of the present invention is: a fire detection method based on an anchor-free network, including the following steps:

[0006] Step 1, Image Acquisition: Obtain a real-time video stream through a camera and convert it into an image format in a computer;

[0007] Step 2, Foreground Judgment: Perform foreground processing on the acquired real-time monitoring image, extract foreground targets from the image, and determine whether there are suspicious moving foreground targets in the image;

[0008] Step 3, Network Detection. If it is determined that a certain frame of image has suspicious moving foreground targets, then send this frame of image into a trained fire detection network to judge the flame targets. First, perform feature extraction through the backbone network, then use the optimized fusion network for feature fusion, and finally predict the flame target category and location information through the classification and regression branches of the detection network. The structure of the fire detection network is as follows:

[0009] The overall form of the fire detection network adopts an anchor-free mechanism, which can greatly adapt to the changing shapes of flames. Based on the FCOS network as the basic network, it is adaptively optimized and improved according to the characteristics of flame targets. The optimization and improvement include: selecting the ShuffleNetV2 lightweight network as the backbone network and embedding a collaborative attention module in each of its stages to enhance the feature extraction ability of the backbone network while maintaining real-time performance; at each stage of the lightweight network ShuffleNetV2 optimized by adding a collaborative attention module, the feature information with different depth levels extracted is sent to the fusion network. The fusion network performs preliminary feature fusion based on the feature pyramid structure and adds a feature enhancement module on the basis of the feature pyramid structure to weight and integrate the feature information at each level, obtain richer feature information and strengthen the original features, and perform feature refinement through the self-attention mechanism module to improve the recognition rate of flame feature information; then, the feature information processed by the above enhancement and refinement is sent into the fire detection network for classification and regression to judge the target category and location information in the image;

[0010] Step 4: Alarm Response; If the prediction result of the network is that there are flame targets in the picture, immediately display the location information of the flame targets and start the alarm program.

[0011] Furthermore, in Step 3, during the training process of the network, a dynamic sampling method is adopted. By combining the foreground probability score and the center prior position weight, the positive sample weight is calculated. By combining the foreground probability score and the center prior position weight, the positive sample weight can be calculated, thereby dynamically guiding the training process of the network.

[0012] Further, in step 2, foreground processing is performed on the acquired real-time monitoring image. The method is as follows: The image is preprocessed using median filtering technology to reduce the influence of irrelevant noise; the ViBe moving foreground extraction algorithm is used to extract foreground targets from the image, and whether there are suspicious moving foreground targets in the image is judged.

[0013] The fire detection network of the present invention abandons the traditional pre-setting of anchor boxes, and adopts an anchor-free mechanism to greatly adapt to the variable scale and shape of flames. Based on the FCOS network, adaptive improvements and optimizations are made for the significant characteristics of flame targets. The backbone network part uses the lightweight network ShuffleNetV2 that takes into account both accuracy and speed, and a collaborative attention module is embedded in each stage, significantly enhancing the feature extraction ability of the backbone network while maintaining real-time performance; the fusion network part adds a feature enhancement module on the basis of the original FPN, and enriches multi-scale feature information through two-level enhancement operations of weighted integration enhancement and feature refinement enhancement, comprehensively improving the detection ability of the network for multi-scale flame targets; during the training process of the network, a dedicated dynamic sampling method is designed according to the variable characteristics of flame shapes, which can screen out high-quality positive samples according to different flame shapes, comprehensively improving the learning effect of the network. The optimized and improved fire detection network has good detection performance and is more suitable for fire detection tasks. The present invention not only has high detection accuracy, has good detection effects on various scale and shape flames under complex environmental backgrounds, but also has strong anti-interference ability, has an obvious inhibitory effect on common false alarm factors, can operate stably under actual monitoring systems, and at the same time has a low model complexity, can achieve real-time detection processes, and has good comprehensive performance and application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a schematic diagram of the overall details of the present invention;

[0015] Figure 2 It is a schematic diagram of the dynamic sampling calculation process in the present invention;

[0016] Figure 3 It is a schematic diagram of the detection process of the present invention;

[0017] Figure 4 It is a schematic diagram of the overall structure of the fire detection network in the present invention;

[0018] Figure 5 It is a schematic diagram of the backbone network structure in the present invention;

[0019] Figure 6 It is a schematic diagram of the structure of the feature enhancement module in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] For further illustration, the specific implementation details of the present invention will be described in detail below in the form of the accompanying drawings, but they should not be construed as limiting the protection scope of the present invention.

[0021] As Figure 1 shown, a fire detection method based on an anchor-free network has the following overall implementation details:

[0022] 1) Construction of the dataset and training of the network model: Since there is currently a lack of high-quality public fire datasets, a fire dataset containing 13,575 pictures has been self-built based on public resources on the Internet and existing resources in the laboratory. This dataset contains rich fire scenes and various flame forms, which can fully meet the needs of actual fire detection tasks.

[0023] On this dataset, the optimized anchor-free fire detection network is trained. During the training process, by combining the foreground probability score and the center prior position weight, the positive sample weight can be calculated, thereby dynamically guiding the training process of the network. The calculation process of the dynamic sampling method is as Figure 2 shown. At the same time, to prevent overfitting, data augmentation methods such as random flipping and brightness transformation are used to process the input pictures.

[0024] After successful training, a dedicated fire detection network model will be obtained, which can be applied to the video surveillance platform for daily real-time detection. Its detection process is as Figure 3 shown.

[0025] 2) Image acquisition: First, the video information stream is read from the video surveillance platform in the form of the main code stream, decoded into the YUV format, then converted into an RGB format image, and normalized to a resolution of 1280*960 to ensure that the clarity of the image can meet the detection requirements.

[0026] 3) Foreground judgment: The obtained video image is sent to the foreground processing module. First, median filtering technology is used for processing to reduce the influence of irrelevant noise information and improve the image quality. Next, the ViBe moving foreground detection algorithm is used to process each frame of the image. First, a background model is established based on the first frame image, and then the foreground area in each frame of the image is determined, and the background models of foreground points and background points are randomly updated. After processing by the foreground module, it can be judged whether there are suspicious foreground moving targets in the image.

[0027] 4) Network detection: If the foreground processing module determines that there are moving foreground targets in the image, then this frame of the image is sent to the trained fire detection network for prediction to determine whether there are flame targets in the image. The overall structure of the fire detection network in the present invention is as Figure 4As shown. After the network receives the input image, it first needs to go through each stage of the lightweight network ShuffleNetV2 optimized by adding a collaborative attention module, extract feature information with different depth levels, and send it to the fusion network. Its backbone network structure is as Figure 5 shown. The fusion network first receives the above feature information and performs preliminary feature fusion based on the feature pyramid structure, and then is enhanced again through the feature enhancement module. The structure of the feature enhancement module is as Figure 6 shown. The specific implementation process is to first weight and integrate the feature information of each level to obtain richer feature information and strengthen the original features. Next, the feature refinement is performed through the self-attention mechanism module to improve the recognition of the flame feature information. Finally, the feature information is sent to the detection network for classification and regression to determine the target category and location information in the image.

[0028] 5) Alarm response: If the fire detection network outputs valid flame target information, it is determined that a fire has occurred in the video surveillance area. At this time, the specific location and situation of the fire are immediately displayed, and the alarm program is started.

Claims

1. A fire detection method based on an anchor-free network, comprising the following steps: Step 1, image acquisition: Obtain a real-time video stream through a camera and convert it into an image format in a computer; Step 2, foreground judgment: Perform foreground processing on the acquired real-time monitoring image, extract foreground objects from the image, and determine whether there are suspicious moving foreground objects in the image; Step 3, network detection. If it is determined that a certain frame of image has suspicious moving foreground objects, then send this frame of image into a trained fire detection network to judge the flame target. First, perform feature extraction through the backbone network, then use the optimized fusion network for feature fusion, and finally predict the flame target category and location information through the classification and regression branches of the detection network; The structure of the fire detection network is as follows: The overall form of the fire detection network adopts an anchor-free mechanism, which can greatly adapt to the changing shapes of flames. Based on the FCOS network as the basic network, and make adaptive optimization improvements according to the characteristics of the flame target. The optimization improvements include: Select the ShuffleNetV2 lightweight network as the backbone network, and embed a collaborative attention module in each stage of it to improve the feature extraction ability of the backbone network while maintaining the real-time effect; After each stage of the lightweight network ShuffleNetV2 optimized by adding the collaborative attention module, the feature information with different depth levels extracted is sent to the fusion network. The fusion network performs preliminary feature fusion based on the feature pyramid structure, and adds a feature enhancement module on the basis of the feature pyramid structure to weight and integrate the feature information of each level, obtain richer feature information and strengthen the original features, and perform feature refinement through the self-attention mechanism module to improve the recognition rate of the flame feature information; Then, the feature information processed by the above enhancement and refinement is sent into the fire detection network for classification and regression to judge the target category and location information in the image; Step 4: Alarm response; If the prediction result of the network is that there is a flame target in the picture, immediately display the location information of the flame target and start the alarm program.

2. The fire detection method according to claim 1, characterized in that, In step 3, during the training process of the network, a dynamic sampling method is adopted. By combining the foreground probability score and the center prior position weight, the positive sample weight is calculated. By combining the foreground probability score and the center prior position weight, the positive sample weight can be calculated, thereby dynamically guiding the training process of the network.

3. The fire detection method according to claim 1, characterized in that, In step 2, when performing foreground processing on the acquired real-time monitoring image, the method is: Use median filtering technology to preprocess the image, to reduce the influence of irrelevant noise; Use the ViBe moving foreground extraction algorithm to extract foreground objects from the image, and determine whether there are suspicious moving foreground objects in the image.

Images