Improved RT-DETR smoke and flame detection method and system

Abstract

The invention discloses an improved RT-DETR smoke and flame detection method and system, and relates to the technical field of flame and smoke detection, and the method comprises the steps: collecting data, constructing a data set, and carrying out the preprocessing; training parameters of the RT-DETR model are set; and improving the RT-DETR model based on the training parameters. According to the improved RT-DETR smoke and flame detection method provided by the invention, training parameters such as different training rounds and batch sizes are set, so that reasonable utilization of training resources and guarantee of model convergence are realized. By improving the RT-DETR model, the effectiveness of feature extraction is improved, and resource consumption caused by too large parameter quantity is avoided. Through the reasonable convolution kernel design and the hierarchical scale feature pyramid network, the accuracy and robustness of detection are improved, so that the beneficial effect of real-time and accurate flame and smoke detection of the model in a complex environment is achieved, and better effects are achieved in the aspects of accuracy, recall rate and average precision.

Description

Technical Field

[0001] The present invention relates to the technical field of flame and smoke detection, and in particular to an improved RT-DETR smoke and flame detection method and system. Background Art

[0002] With the continuous advancement of urbanization and industrialization, the prevention and control of fire hazards has gradually become an important public safety task. As a key component of the fire warning system, flame and smoke detection technology can achieve rapid response to the early stages of fire. Traditional fire detection methods are mainly based on temperature, gas and optical sensors, but due to their strong dependence on the environment, there are problems such as untimely detection and high false alarm rate. In recent years, with the development of computer vision technology, flame and smoke detection technology based on deep learning has received widespread attention. In particular, real-time target detection models (such as RT-DETR) have gradually been applied to flame and smoke detection tasks. Their efficient detection speed and high accuracy meet the needs of early fire warning. However, traditional detection methods and initially applied deep learning models still have many shortcomings when dealing with complex backgrounds and multi-scale flame and smoke characteristics, and cannot fully cope with diverse fire scenes.

[0003] Although the current flame and smoke detection methods based on deep learning have improved in accuracy, they still have great limitations in complex environments and multi-scale feature detection. First, the basic block design of the existing RT-DETR model shows the problem of large computational complexity and large number of parameters when dealing with the sparsity and complex background of flames and smoke, and it is difficult to efficiently deploy on edge devices with limited resources. Although directly introducing large convolution kernels can enhance the feature extraction capability of the model, it will significantly increase the consumption of computing resources, greatly reduce the training and reasoning efficiency of the model, and it is difficult to adapt to the requirements of real-time detection. Secondly, in terms of cross-scale feature fusion, the cross-scale feature fusion module (CCFM) used in RT-DETR has limited integration effect on features of different scales, especially in capturing the subtle features of flames and smoke. The morphology of flames and smoke varies with the environment. The existing methods lack hierarchical processing of multi-scale feature maps, resulting in the inability of the model to stably capture key features at different viewing angles and distances, affecting the detection accuracy. In addition, the traditional feature fusion strategy simply superimposes features of different layers, fails to effectively balance the combination of high-level semantic information and low-level positioning accuracy, and easily leads to feature fuzziness. In contrast, the HS-FPN structure proposed by our invention has realized the hierarchical screening and integration of multi-scale features through the improvement of feature selection and feature fusion modules. In particular, the CA module compresses redundant information while retaining feature details, thereby significantly improving the accuracy and robustness of detection. At the same time, by improving the convolution kernel design and adopting sparse group convolution, the calculation amount of the model is effectively reduced, making it more real-time and accurate in complex detection scenarios of flames and smoke. Summary of the invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is: the existing flame and smoke detection methods have insufficient detection accuracy, poor cross-scale feature fusion effect, and difficulty in efficient deployment, as well as the problem of how to accurately detect flame and smoke characteristics in real time in a complex environment.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: a smoke and flame detection method for improving RT-DETR, comprising collecting data to construct a data set for preprocessing; setting RT-DETR model training parameters; and improving the RT-DETR model based on the training parameters.

[0007] As a preferred solution of the improved RT-DETR smoke and flame detection method described in the present invention, the data acquisition and data set construction for preprocessing includes acquiring flame and smoke images and constructing a flame and smoke data set.

[0008] As a preferred solution of the smoke and flame detection method of the improved RT-DETR of the present invention, wherein: the preprocessing of the collected data to construct a data set includes preprocessing the collected flame and smoke images, adjusting different image sizes to be standardized to the model input size;

[0009] Through data enhancement, random cropping, rotation, and color change are performed;

[0010] Use the image processing library functions in 0penCv to rotate the image by 90°, blur it, change its color, and add noise.

[0011] The dataset is divided into training set, validation set and test set in a ratio of 7:2:1.

[0012] As a preferred solution of the improved RT-DETR smoke and flame detection method of the present invention, wherein: the setting of the RT-DETR model training parameters includes determining the rounds of model convergence by setting different training rounds;

[0013] Set the training parameters of the original RT-DETR model, including:

[0014] Training rounds, batch size, number of data loading threads.

[0015] As a preferred solution of the smoke and flame detection method of the improved RT-DETR of the present invention, wherein: the setting of the RT-DETR model training parameters includes introducing a shift operator into the original RT-DETR;

[0016] Transform the convolution kernel and apply a shift operation to each convolution result;

[0017] The transformed convolution kernel is expressed as:

[0018]

[0019] Among them, (i, j) represents the position of the sliding window on the image or feature map, k represents the number of N×N convolution kernels decomposed into an M×N convolution kernel, kw and kh represent the dimensions of the convolution kernel, Indicates rounding up, p represents position information, w and x represent the weights and features of the corresponding positions, Δm and Δn represent the offsets used to calculate the current sliding window position in the convolution operation, and Δp represents the offset associated with the decomposition of the convolution kernel.

[0020] As a preferred solution of the smoke and flame detection method of improving RT-DETR of the present invention, wherein: the improved RT-DETR model based on training parameters includes redesigning a hierarchical scale-based feature pyramid network, including feature selection and feature fusion;

[0021] Feature selection includes a channel attention module processing the input feature map, expressed as:

[0022] f∈R C×H×W

[0023] Among them, C represents the number of channels, H and W represent the height and width of the feature map respectively.

[0024] As a preferred solution of the smoke and flame detection method of improving RT-DETR described in the present invention, wherein: the improved RT-DETR model based on training parameters includes a double pooling process of global average pooling and global maximum pooling of the input feature map, and the output features are fused;

[0025] Different dimensions need to be matched before feature fusion;

[0026] Feature fusion involves processing the high-level features of the input using inverse convolution, and then upsampling or downsampling the processed high-level features through bilinear interpolation to match the size of the low-level features;

[0027] Convert high-level features into attention weights, filter low-level features with weights, and fuse the filtered low-level features with high-level features.

[0028] Another object of the present invention is to provide an improved RT-DETR smoke and flame detection system, which can solve the problem of insufficient detection accuracy of current flame and smoke detection methods by improving the RT-DETR model based on training parameters.

[0029] As a preferred solution of the improved RT-DETR smoke and flame detection system described in the present invention, it includes: an initialization module, a model training parameter setting module, and a model improvement module; the initialization module is used to collect flame and smoke images, construct flame and smoke data sets, and perform data preprocessing and division; the model training parameter setting module is used to set RT-DETR model training parameters; the model improvement module is used to improve the RT-DETR model.

[0030] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of improving the smoke and flame detection method of RT-DETR.

[0031] A computer readable storage medium stores a computer program, which, when executed by a processor, implements the steps of improving the smoke and flame detection method of RT-DETR.

[0032] Beneficial effects of the present invention: The smoke and flame detection method of the improved RT-DETR provided by the present invention improves the detection accuracy and stability of the model by preprocessing the collected flame and smoke images. By setting different training parameters such as training rounds and batch sizes, the convergence of the model under specific training conditions is determined, and the rational use of training resources and the guarantee of model convergence are achieved. By finely setting the training parameters, it is ensured that the model achieves optimal performance within appropriate rounds, so that the training process is carried out efficiently without wasting computing power. Through reasonable parameter configuration, the model can converge stably during the training process, avoiding overfitting or underfitting, thereby ultimately achieving the beneficial effect of improving the stability and performance of the model. By improving the RT-DETR model, the simulation of the effect of large convolution kernels is realized, which not only increases the effectiveness of feature extraction, but also avoids resource consumption caused by excessive parameters. In addition, through the HS-FPN structure, features of different scales are screened and fused, and the feature selection module (CA module) and the dimension matching module are combined to optimize the combination of high-level and low-level features. Through reasonable convolution kernel design and hierarchical scale feature pyramid network, the model can capture the key features of flames and smoke in different scenes and scales, improve the accuracy and robustness of detection, and thus achieve the beneficial effect of real-time and accurate detection of flames and smoke in complex environments. The present invention achieves better results in accuracy, recall rate and average precision. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0034] Figure 1 An overall flow chart of an improved RT-DETR smoke and flame detection method provided for the first embodiment of the present invention.

[0035] Figure 2 A shift operator model diagram of an improved RT-DETR smoke and flame detection method provided by the second embodiment of the present invention.

[0036] Figure 3 A hierarchical-scale feature pyramid network diagram of an improved RT-DETR smoke and flame detection method provided in the second embodiment of the present invention.

[0037] Figure 4The mAP50 graph of the original RT-DETR on the smoke and flame dataset of an improved RT-DETR smoke and flame detection method provided by the second embodiment of the present invention.

[0038] Figure 5 An mAP50 graph of an improved RT-DETR on a smoke and flame dataset of a smoke and flame detection method based on an improved RT-DETR provided in the second embodiment of the present invention.

[0039] Figure 6 An overall flow chart of an improved RT-DETR smoke and flame detection system provided in the third embodiment of the present invention. DETAILED DESCRIPTION

[0040] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0041] Example 1, reference Figure 1-Figure 5 , as an embodiment of the present invention, provides a smoke and flame detection method of improving RT-DETR, comprising:

[0042] S1: Collect data and construct a data set for preprocessing.

[0043] Furthermore, we collect flame and smoke images in urban buildings, industrial environments, forests, etc. to build a diverse flame and smoke dataset.

[0044] It should be noted that the collected flame and smoke images were preprocessed, including adjusting different image sizes to the standardized model input size of 640*640. Data diversity was increased through data enhancement, such as random cropping, rotation, color transformation, etc. The specific method is: using the image processing library function in OpenCv, the image was rotated 90°, blurred, color transformed, and noise added.

[0045] The dataset is divided into training set, validation set and test set in a ratio of 7:2:1 to ensure that the categories in each subset are evenly distributed.

[0046] S2: Set RT-DETR model training parameters.

[0047] Furthermore, by setting different training rounds, the specific rounds at which the model begins to converge are determined. After experiments, the main training parameters of the original RT-DETR model are set, including: setting the training rounds to 200 rounds to ensure the convergence of the model, setting the batch size to 8 to ensure that it does not exceed the maximum memory capacity of the GPU, setting the number of data loading threads to 4, and keeping other parameters consistent by default.

[0048] S3: Improve the RT-DETR model based on training parameters.

[0049] Furthermore, the original RT-DETR flame and smoke detection model was improved. The specific measures include: 1. In the backbone of the original RT-DETR, the original basic blocks were improved, and a more efficient convolution kernel re-parameterization technology was introduced to optimize the convolution operation, reducing the overall computational complexity and parameter amount of the model.

[0050] The backbone network in the original RT-DETR model uses a standard 3*3 convolution kernel. Although large convolution kernels (such as 5*5, 11*11 convolution kernels) have advantages in feature extraction and organization, directly introducing large convolution kernels will cause the number of model parameters to surge, making the model training extremely difficult and difficult to deploy on edge devices. To solve this problem, a method is proposed: achieve the effect of a large convolution kernel by translating the standard convolution, and then obtain a sparse group convolution by pruning. The above operation is referred to as the shift operator. By introducing the shift operator, more features can be extracted, thereby improving the accuracy of flame and smoke detection. At the same time, this method can establish long-term sparse dependencies, which is very important for flame and smoke detection.

[0051] The shift operator (shift) is introduced in the basic block of the original RT-DETR backbone. Specifically, the large convolution kernel is converted into a series of normal smaller convolution kernels, and then a shift operation is applied to each convolution result. The improved module structure is as follows Figure 1 As shown in the figure, a 15x3 convolution is decomposed into 5 3x3 convolution kernels and a shift operation is applied.

[0052] Figure 1 The formula is needed to decompose the large convolution kernel into several small convolution kernels:

[0053]

[0054] Among them, (i, j) represents the position of the sliding window on the image or feature map, k is the number of N×N convolution kernels decomposed into an M×N convolution kernel, kw and kh are the dimensions of the convolution kernel, Indicates rounding up, and p indicates position information. w and x are the weights and features of the corresponding positions, Δm and Δn are used in the convolution operation to calculate the offset relative to the current sliding window position, and Δp is the offset related to the decomposition of the convolution kernel.

[0055] It should be noted that the original cross-scale feature fusion module (CCFM) was improved by using the hierarchical scale-based feature pyramid network (HS-FPN) to complete multi-scale feature fusion. This enables the model to capture more comprehensive flame and smoke feature information.

[0056] Since the size of flames and smoke can vary greatly depending on the distance and viewing angle, even the size of the same flame or smoke can be different under different conditions. To solve this problem, a hierarchical scale-based feature pyramid network (HS-FPN) structure is introduced. HS-FPN mainly consists of two parts: (1) feature selection module. (2) feature fusion module. First, feature maps of different scales are screened in the feature selection module. Subsequently, the high-level and low-level information in the feature map are fused through the selective feature fusion (SFF) mechanism. This fusion produces semantically rich features that help detect subtle features in flame and smoke images, thereby enhancing the detection ability of the model.

[0057] The feature selection process is mainly completed by the channel attention (CA) module and the dimension matching module. The channel attention module processes the input feature map, which is expressed as:

[0058] f∈R C×H×W

[0059] Where C represents the number of channels, and H and W represent the height and width of the feature map, respectively. The input feature map needs to go through a double pooling process of global average pooling and global maximum pooling, and then the resulting features are combined. Next, the weight value of each channel is calculated by the Sigmoid activation function. Pooling has three main purposes: downsampling the feature map to reduce the dimension; eliminating redundant data to compress the features and reduce the number of parameters; and achieving invariance to translation, rotation, and scaling. The combination of these pooling methods in the CA module helps to extract the most representative information from each channel while minimizing information loss. The weighted information is then multiplied with the feature map of the corresponding scale to generate a filtered feature map. Before feature fusion, features of different dimensions need to be dimensionally matched because they have different numbers of channels. To this end, the dimension matching module uses 1x1 convolution to set the number of output channels of feature maps of different dimensions to 256.

[0060] When processing the flame and smoke detection dataset, the SFF module can optimize the combination of high-level and low-level features. Specifically, the multi-scale feature maps generated by the backbone network perform well in terms of semantic information, but are relatively rough in terms of target localization. In contrast, low-level features can accurately locate the target, but their semantic information is relatively scarce, and the SFF module can solve this problem well. Instead of simply adding the pixel values ​​of high-level features and low-level features, this module adopts a strategic feature fusion method. It uses high-level features as weights to selectively filter and fuse basic semantic information in low-level features. The specific steps are as follows: First, the input high-level features are processed using inverse convolution (T-Conv), and then the processed high-level features are upsampled or downsampled by bilinear interpolation to match the size of the low-level features. Next, the CA module is used to convert the high-level features into corresponding attention weights, and the low-level features are filtered with these weights. Finally. The filtered low-level features are fused with the high-level features to enhance the feature representation ability of the model on the flame and smoke detection dataset. In this way, the model can effectively combine the precise positioning ability of low-level features with the rich semantic information of high-level features, thereby achieving better performance in fire and smoke detection tasks.

[0061] Going a step further, we introduce the shift operator:

[0062] Implementation conditions: Introduce shift operators into the basic blocks of the original RT-DETR backbone and then replace the original basic blocks.

[0063] Parameters: The original 3*3 convolution kernel can now be flexibly adjusted to a 7*7 or 9*9 large convolution kernel.

[0064] Redesign the original Cross-Scale Feature Fusion Module (CCFM).

[0065] Implementation conditions: The feature selection module and the feature fusion module are connected according to Figure 2 The order of design is to form a hierarchical scale-based feature pyramid network (HS-FPN)

[0066] Parameters: The parameter settings of the network structure, such as the number of layers, the number of channels in each layer, etc., need to be adjusted according to the actual situation.

[0067] It should be noted that the area under the PR curve is defined as: Where AP is the average precision, P and R represent precision and recall respectively, and P(R) represents the precision function. In order to solve the single point value limitation of P and R, this paper will use mAP50 as the verification indicator, and the calculation formula is:

[0068]

[0069] Where N is the total number of categories, AP i is the AP value of the ith category. The larger the area under the PR curve, the better. Therefore, the mAP50 of the original RT-DETR on the smoke and fire dataset is as follows: Figure 4 As shown in the figure, the mAP50 of the model improved by the method is as follows Figure 5 shown.

[0070] Example 2, an embodiment of the present invention, provides an improved RT-DETR smoke and flame detection method. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0071] First, the collected flame and smoke datasets are input into the improved RT-DETR flame and smoke detection model for model training. The training summarizes the results of four ablation experiments. They are the original RT-DETR model experiment, the RT-DETR+shift operator model experiment, the RT-DETR+HS-FPN model experiment, and the model experiment after the overall improvement of RT-DETR. From the ablation experiment, it can be seen that the improved model has an accuracy (Precision) of 1.2%, a recall rate (Recal1) of 2.9%, and a mAP of 1.7% compared with the original model, and the number of model parameters has been reduced by 29%.

[0072] Table 1 Ablation experiment results

[0073]

[0074] The performance of the improved RT-DETR flame and smoke detection model was evaluated, and a comparative experiment was conducted with other models with better performance. From the comparative experiments, it can be seen that the accuracy, recall rate and average precision have been significantly improved, which further demonstrates that the present invention is superior to the prior art.

[0075] Table 2 Comparative test results

[0076]

[0077] Example 3, reference Figure 6 , as an embodiment of the present invention, provides an improved RT-DETR smoke and flame detection system, including an initialization module, a model training parameter setting module, and a model improvement module.

[0078] The initialization module is used to collect flame and smoke images, construct flame and smoke datasets and perform data preprocessing and division. The model training parameter setting module is used to set the RT-DETR model training parameters. The model improvement module is used to improve the RT-DETR model.

[0079] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0080] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0081] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0082] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc. It should be noted that the above embodiments are only used to illustrate the technical solution of the present invention and are not limited. Although the present invention is described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

[0083] It should be noted that the above embodiments are only used to illustrate the technical solutions 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 the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A smoke and flame detection method for improving RT-DETR, characterized in that: include: Collect data to build a data set for preprocessing; Set RT-DETR model training parameters; Improve the RT-DETR model based on training parameters.

2. The improved RT-DETR smoke and flame detection method as claimed in claim 1, characterized in that: The collecting data and constructing a data set for preprocessing includes collecting flame and smoke images and constructing a flame and smoke data set.

3. The improved RT-DETR smoke and flame detection method as claimed in claim 2, characterized in that: The collected data is constructed into a data set for preprocessing, including preprocessing the collected flame and smoke images, adjusting different image sizes to be standardized to the model input size; Through data enhancement, random cropping, rotation, and color change are performed; Use the image processing library functions in 0penCv to rotate the image by 90°, blur it, change its color, and add noise. The dataset is divided into training set, validation set and test set in a ratio of 7:2:

1.

4. The improved RT-DETR smoke and flame detection method as claimed in claim 3, characterized in that: The setting of the RT-DETR model training parameters includes determining the rounds of model convergence by setting different training rounds; Set the training parameters of the original RT-DETR model, including: Training rounds, batch size, number of data loading threads.

5. The improved RT-DETR smoke and flame detection method as claimed in claim 4, characterized in that: The setting of the RT-DETR model training parameters includes introducing a shift operator into the original RT-DETR; Transform the convolution kernel and apply a shift operation to each convolution result; The transformed convolution kernel is expressed as: Among them, (i, j) represents the position of the sliding window on the image or feature map, k represents the number of N×N convolution kernels decomposed into an M×N convolution kernel, kw and kh represent the dimensions of the convolution kernel, Indicates rounding up, p represents position information, w and x represent the weights and features of the corresponding positions, Δm and Δn represent the offsets used to calculate the current sliding window position in the convolution operation, and Δp represents the offset associated with the decomposition of the convolution kernel.

6. The improved RT-DETR smoke and flame detection method as claimed in claim 5, characterized in that: The improved RT-DETR model based on training parameters includes redesigning a hierarchical scale-based feature pyramid network, including feature selection and feature fusion; Feature selection includes a channel attention module processing the input feature map, expressed as: f∈R C×H×W Among them, C represents the number of channels, H and W represent the height and width of the feature map respectively.

7. The improved RT-DETR smoke and flame detection method as claimed in claim 6, characterized in that: The improved RT-DETR model based on training parameters includes a double pooling process of global average pooling and global maximum pooling of the input feature map to fuse the output features; Different dimensions need to be matched before feature fusion; Feature fusion involves processing the high-level features of the input using inverse convolution, and then upsampling or downsampling the processed high-level features through bilinear interpolation to match the size of the low-level features; Convert high-level features into attention weights, filter low-level features with weights, and fuse the filtered low-level features with high-level features.

8. A system using the improved RT-DETR smoke and flame detection method as claimed in any one of claims 1 to 7, characterized in that: Including initialization module, model training parameter setting module, model improvement module; The initialization module is used to collect flame and smoke images, construct flame and smoke data sets, and perform data preprocessing and division; The model training parameter setting module is used to set the RT-DETR model training parameters; The model improvement module is used to improve the RT-DETR model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the smoke and flame detection method for improving RT-DETR according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the smoke and flame detection method for improving RT-DETR according to any one of claims 1 to 7 are implemented.

Images