Outdoor scene smoke area identification method based on deep learning

Abstract

The invention provides an outdoor scene smog area identification method based on deep learning. The method comprises the steps of S01, obtaining an image data set containing different outdoor scene images to construct a training set and a test set; step S02, constructing and training a model; an RDAACN deep learning model is constructed, the model comprises encoders of a plurality of convolution blocks connected in sequence and a decoder for outputting a prediction result according to image features extracted by the encoders, and the model is trained; and S03, testing the trained RDAACN deep learning model by using a test set to identify a smoke region in the image. The method has the advantages of being simple in principle and high in intelligent degree, the pixel-level precision extraction result of the smoke area in the image can be output, and the efficiency and accuracy of identifying the smoke area in the outdoor scene image are improved.

Description

Technical Field

[0001] The present invention relates to the field of computer vision and deep learning technology, and in particular to a method for identifying smoke areas in outdoor scenes based on deep learning. Background Art

[0002] Fire is a sudden disaster, so its early monitoring and early warning are particularly urgent. The smoke generated in the early stage of a fire is an extremely critical early warning signal. Whether the smoke area can be quickly and accurately identified is directly related to whether effective prevention and control measures can be taken at the beginning of the fire, thereby minimizing the losses. In the process of smart city construction, refined management of urban disasters has become a key link, and its core lies in strengthening the safety risk monitoring and early warning system with the help of cutting-edge technology. For the early monitoring of fires, highly accurate smoke area identification technology is indispensable.

[0003] In the past, computer vision algorithms have dominated the field of smoke recognition. Its operating mechanism is to first extract the color, texture and other features of the smoke, and then use the classification model to make judgments. However, in actual operation, this method has exposed many defects. The translucent nature of smoke makes it easy to blend with the complex background of the outdoors. The complex and changeable lighting conditions and meteorological factors further increase the difficulty of recognition, making it difficult for the recognition accuracy to meet the expected standards. At the same time, the cumbersome algorithm process leads to a high false alarm rate, which greatly weakens its practicality in disaster management practice.

[0004] With the vigorous development of deep learning technology, the field of semantic segmentation has achieved remarkable innovative results. The emergence of advanced models such as fully convolutional networks (FCN), U-Net, and SegNet, with end-to-end learning architecture, has achieved accurate pixel-level classification of images, achieved a qualitative leap in recognition accuracy and processing efficiency, and has been successfully applied in many fields such as medical image analysis and autonomous driving scene perception. However, in the specific and challenging task of smoke area recognition in outdoor scenes, related research and application are still in their infancy. The high complexity of the outdoor environment, including dynamic changes in light intensity and angle, complex and diverse background objects, and irregular evolution of smoke morphology, poses a severe test to the generalization performance and robustness of the model. The existing deep learning models have not been fully optimized and adapted to deal with such special scenarios, and it is difficult to effectively meet the urgent needs of smoke area recognition for early outdoor fire monitoring. Summary of the invention

[0005] The technical problem to be solved by the present invention is: in response to the technical problems existing in the prior art, the present invention provides an outdoor scene smoke area recognition method based on deep learning with simple principle, high intelligence, high recognition efficiency and accuracy.

[0006] In order to solve the above technical problems, the technical solution proposed by the present invention is: A method for identifying smoke areas in outdoor scenes based on deep learning, comprising the following steps: Step S01: Acquisition of outdoor scene data set; Obtain an image dataset containing images of different outdoor scenes and construct a training set and a test set; Step S02: Model construction and training; Constructing an RDAACN deep learning model, the RDAACN deep learning model includes an encoder and a decoder; the encoder includes a plurality of sequentially connected convolution blocks RDAACB, the convolution block RDAACB includes a plurality of asymmetric convolution layers with different dilation rates, for extracting image features; the decoder is used to output a prediction result according to the image features extracted by the encoder, the asymmetric convolution layer means that the convolution kernel has different lengths and widths; Using the training set and the test set to train the constructed RDAACN deep learning model, and saving the trained RDAACN deep learning model; Step S03: inputting the outdoor scene image to be identified into the trained RDAACN deep learning model to identify the smoke area in the outdoor scene image to be identified; The plurality of sequentially connected convolution blocks RDAACB in the encoder include sequentially connected convolution blocks RDAACB1, convolution blocks RDAACB2, convolution blocks RDAACB3 and convolution blocks RDAACB4; The parameter calculation formula of the convolution block RDAACB4 is: , Among them, Parameters represents the parameter quantity, F CH represents the height of the convolution kernel, F CW represents the width of the convolution kernel, C I Indicates the number of channels of the input feature map, C O Indicates the number of channels of the output feature map.

[0007] As a further improvement of the method of the present invention: in the step S02, the convolution block RDAACB includes a first branch for generating a first feature map and a second branch for generating a second feature map, and the first feature map and the second feature map are added to obtain an output feature map.

[0008] As a further improvement of the method of the present invention: the method for obtaining the output feature map includes: Obtain the original feature map or input image, and input the original feature map or input image into the first branch. The number of convolution kernels is N e / 2, the expansion rate is 3, the output is processed by batch normalization, the activation function is the 1×3 hole convolution layer of the rectified linear unit ReLU, and the output feature map of the 1×3 hole convolution layer is input to the number of convolution kernels. N e / 2, the expansion rate is 3, the output is processed by batch normalization, the activation function is the 3×1 hole convolution layer with the rectified linear unit ReLU, and the output feature map of the 3×1 hole convolution layer is input to the number of convolution kernels. N e / 2, the expansion rate is 6, the output is processed using batch normalization, and the activation function is a 3×3 hollow convolution layer with a rectified linear unit ReLU; The output feature maps of the 3×1 dilated convolution layer and the 3×3 dilated convolution layer are fused by channel merging, and the fused feature maps are input into the convolution kernel with the number of N e , the activation function uses a 1×1 convolution layer of Sigmoid, and the output feature map of the 1×1 convolution layer is multiplied by the feature map obtained by feature fusion to obtain a first feature map; The original feature map or input image is input into the second branch. The number of convolution kernels is N e The output is processed by batch normalization, and the activation function is a 1×3 convolution layer with a rectified linear unit ReLU. The output feature map of the 1×3 convolution layer is then input to the convolution kernel with the number of N e , the output is processed by batch normalization, and the activation function uses a 3×1 convolution layer with a rectified linear unit ReLU to obtain the second feature map; The first feature map is added to the second feature map to obtain an output feature map of the convolution block RDAACB.

[0009] As a further improvement of the method of the present invention: the output ends of the convolution blocks RDAACB1, RDAACB2 and RDAACB3 are also connected to a maximum pooling layer with a window size of 2×2 and a pooling window step size of 2 for outputting the output feature map after maximum pooling, and the output end of the convolution block RDAACB4 does not include a maximum pooling layer but directly outputs the output feature map.

[0010] As a further improvement of the method of the present invention: the calculation formula of the parameter amount of the convolution block RDAACB1 to the convolution block RDAACB3 is: , Among them, Parameters represents the parameter quantity, F CH represents the height of the convolution kernel, F CW represents the width of the convolution kernel, C I Indicates the number of channels of the input feature map, C O Indicates the number of channels of the output feature map.

[0011] As a further improvement of the method of the present invention: in the step S02, the decoder sequentially connects a 3×3 convolution layer with 512 convolution kernels and a rectified linear unit ReLU activation function, a first upsampling module, a 3×3 convolution layer with 64 convolution kernels and a rectified linear unit ReLU activation function, a second upsampling module, a 3×3 convolution layer with 32 convolution kernels and a rectified linear unit ReLU activation function, and a 1×1 convolution layer with 2 convolution kernels.

[0012] As a further improvement of the method of the present invention: the method in which the decoder outputs the prediction result according to the image features extracted by the encoder is: Step S211: After the feature map output by the convolution block RDAACB4 is input into the 3×3 convolution layer of the decoder, whose number of convolution kernels is 512 and whose activation function is the rectified linear unit ReLU, the output feature map of the 3×3 convolution layer is upsampled by 2 times, and then input into the 3×3 convolution layer, whose number of convolution kernels is 64 and whose activation function is the rectified linear unit ReLU, the obtained output feature map is added to the output feature map of the convolution block RDAACB2, and the feature fusion map is obtained, and then upsampled by 4 times to obtain the sampled feature map; Step S212: The feature fusion image after 4 times upsampling is input into the 3×3 convolution layer with 32 convolution kernels and the activation function of the rectified linear unit ReLU, and then input into the 1×1 convolution layer with 2 convolution kernels to obtain the output prediction of the RDAACN deep learning model.

[0013] As a further improvement of the method of the present invention: the parameter quantity of the RDAACN deep learning model is the sum of the parameter quantities of all convolution blocks.

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention improves the RDAACN deep learning model by using dense dilated convolution and asymmetric convolution kernel convolution block RDAACB structure and connecting multiple convolution blocks RDAACB in series, thereby realizing multi-level and multi-angle extraction of image features, better capturing feature information of different scales and abstraction levels, adapting to complex situations such as variable outdoor smoke forms and cluttered backgrounds, and significantly improving the accuracy and efficiency of smoke area recognition. It not only expands the receptive field and enhances the ability to capture smoke features, but also greatly reduces the number of model parameters, making the model more lightweight and more computationally efficient.

[0015] 2. The present invention further improves the feature extraction capability and robustness of the model through multi-scale feature fusion and residual connection, ensuring good generalization ability in different outdoor scenes. The whole method is simple in principle and efficient in workflow. It can be highly automated from data set acquisition to model training and testing, greatly improving the feasibility and efficiency of practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flow chart of a method for identifying smoke areas in outdoor scenes based on deep learning in an embodiment of the present invention.

[0017] Figure 2 is an example diagram of an image using a data set in an embodiment of the present invention.

[0018] Figure 3 It is a schematic diagram of the structural principle of RDAACB in an embodiment of the present invention.

[0019] Figure 4 It is a schematic diagram of the structural principle of the RDAACN model in a specific application example of the present invention. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] like Figure 1 As shown, the outdoor scene smoke area recognition method based on deep learning in this embodiment includes the following steps: Step S01: Acquisition of outdoor scene data set; Obtain an image dataset containing images of different outdoor scenes and construct training and test sets.

[0022] In this embodiment, the image dataset used is the deepsmoke dataset, which is mainly used in the field of target detection and covers 252 images of rich and diverse outdoor scenes. Each image is equipped with corresponding smoke area annotation information. Some image examples are shown as follows: Figure 2 As shown in (a) in the figure, the corresponding annotations are as follows Figure 2 In the specific operation, 189 images were carefully selected from the 252 images as the training set to train the model so that it can learn the characteristics and laws of the smoke area; the remaining 63 images were set as the test set to test the performance of the trained model in identifying the smoke area.

[0023] Step S02: Model construction and training; Construct an RDAACN deep learning model, which includes an encoder and a decoder; the encoder includes multiple sequentially connected convolution blocks RDAACB, and the convolution block RDAACB includes multiple asymmetric convolution layers with different dilation rates for extracting image features; the decoder is used to output prediction results based on the image features extracted by the encoder, and the asymmetric convolution layer refers to a convolution kernel with different lengths and widths; Use the training set and test set to train the constructed RDAACN deep learning model, and save the trained RDAACN deep learning model.

[0024] In this embodiment, the convolution block RDAACB includes a first branch for generating a first feature map and a second branch for generating a second feature map, and the first feature map and the second feature map are added to obtain an output feature map.

[0025] In this embodiment, the method for obtaining the output feature map includes: Get the original feature map or input image, and input the original feature map or input image into the first branch. The number of convolution kernels is N e / 2, the expansion rate is 3, the output is processed by batch normalization, and the activation function is the 1×3 hole convolution layer with the rectified linear unit ReLU; the output feature map of the 1×3 hole convolution layer is input to the number of convolution kernels N e / 2, expansion rate is 3, output uses batch normalization, activation function uses rectified linear unit ReLU 3×1 hole convolution layer; the output feature map of 3×1 hole convolution layer has the number of input convolution kernels is N e / 2, the expansion rate is 6, the output is processed using batch normalization, and the activation function is a 3×3 hollow convolution layer with a rectified linear unit ReLU; The output feature maps of the 3×1 dilated convolution layer and the 3×3 dilated convolution layer are fused by channel merging, and the fused feature maps are input into the convolution kernel with the number of N e , the activation function uses a 1×1 convolution layer of Sigmoid, and the output feature map of the 1×1 convolution layer is multiplied by the fused feature map to obtain a first feature map; The original feature map or input image is input into the second branch. The number of convolution kernels is N e , the output is processed by batch normalization, the activation function is a 1×3 convolution layer with a rectified linear unit ReLU, and the output feature map of the 1×3 convolution layer is then input to the convolution kernel with the number of N e , the output is processed by batch normalization, and the activation function is a 3×1 convolution layer with a rectified linear unit ReLU to obtain the second feature map; The first feature map is added to the second feature map to obtain the output feature map of the convolution block RDAACB.

[0026] Specifically, the encoder of the RDAACN deep learning model includes 4 convolution blocks RDAACB connected in sequence. It is understandable that the specific number of convolution blocks RDAACB in the encoder can be adjusted according to actual needs. In this embodiment, the multiple sequentially connected convolution blocks RDAACB in the encoder include sequentially connected convolution blocks RDAACB1, convolution blocks RDAACB2, convolution blocks RDAACB3 and convolution blocks RDAACB4, and the output ends of the convolution blocks RDAACB1, RDAACB2 and RDAACB3 are also connected to a maximum pooling layer with a window size of 2×2 and a step size of 2 for pooling the output feature map after maximum pooling, and the output end of the convolution block RDAACB4 does not include a maximum pooling layer but directly outputs the output feature map.

[0027] In this embodiment, the parameter calculation formula of the convolution block RDAACB4 is: , Among them, Parameters represents the parameter quantity, F CH represents the height of the convolution kernel, F CW represents the width of the convolution kernel, C I Indicates the number of channels of the input feature map, C O Indicates the number of channels of the output feature map.

[0028] The calculation formula for the parameter amount of convolution block RDAACB1~convolution block RDAACB3 is: , Among them, Parameters represents the parameter quantity, F CH represents the height of the convolution kernel, F CW represents the width of the convolution kernel, C I Indicates the number of channels of the input feature map, C O Indicates the number of channels of the output feature map.

[0029] In this embodiment, the parameter quantity of the RDAACN deep learning model is the sum of the parameter quantities of all convolution blocks.

[0030] This embodiment uses asymmetric convolution kernels and specific convolution layer connection methods (such as residual connections) to effectively reduce the number of parameters and calculations while ensuring the effect of feature extraction, which helps to improve the efficiency of model training and reasoning. Specifically, in order to retain the image features ignored by the hole convolution, the convolution block RDAACB is based on the original convolution block RDAACB. The feature map input to it is improved from being processed by a 3×3 convolution layer to being processed by a 1×3 asymmetric convolution and a 3×1 asymmetric convolution that can reduce the amount of calculation, and then added using a residual connection. This residual connection method better retains image features that may be ignored by the hole convolution, avoids information loss, and improves the performance and accuracy of the model.

[0031] like Figure 3 As shown in (a), in a specific application embodiment, the structures of RDAACB1 to RDAACB3 include: 1. The original feature map is input into a 1×3 dilated convolution layer with Ne / 2 convolution kernels, dilation rate d of 3, batch normalization (BN) processing on the output, and rectified linear unit ReLU as the activation function. The output feature map of the 1×3 dilated convolution layer is input into a 3×1 dilated convolution layer with Ne / 2 convolution kernels, dilation rate d of 3, BN processing on the output, and ReLU as the activation function. The output feature map of the 3×1 dilated convolution layer is input into a 3×3 dilated convolution layer with Ne / 2 convolution kernels, dilation rate d of 6, BN processing on the output, and ReLU as the activation function. 2. The output feature maps of the 3×1 dilated convolution layer and the 3×3 dilated convolution layer are fused in a channel-merging manner. The fused feature maps are input into a 1×1 convolution layer with Ne convolution kernels and Sigmoid activation function. The output feature map of the 1×1 convolution layer is then multiplied by the feature map obtained by feature fusion. 3. The original feature map is input into a 1×3 convolution layer with Ne convolution kernels, BN processing for output and ReLU activation function. The output feature map of the 1×3 convolution layer is input into a 3×1 convolution layer with Ne convolution kernels, BN processing for output and ReLU activation function. The output feature map of the 3×1 convolution layer is added to the feature map obtained by multiplication in step 2. 4. The added feature map is input into a maximum pooling layer with a pooling window of 2×2 and a step size Sp of 2. The size of the feature map is reduced to 0.5 times of the original feature map to obtain the output feature map of RDAACB.

[0032] like Figure 3 As shown in (b), convolution block RDAACB4 does not include a maximum pooling layer compared to convolution blocks RDAACB1 to RDAACB3.

[0033] In this embodiment, the convolution block RDAACB has dense hole convolutions, which provide a large receptive field and dense pixel sampling, help capture the global information and detail information of the image, and enhance the model's understanding and processing capabilities of complex scenes.

[0034] This embodiment realizes multi-level and multi-angle extraction of image features by connecting multiple convolution blocks RDAACB in series and processing the first branch and the second branch in each convolution block differently. It combines the advantages of different convolution layers and hole convolution layers and can better capture feature information of different scales and abstraction levels.

[0035] In this embodiment, all original convolution blocks RDAACB in the RDAACN deep learning model are replaced with convolution blocks RDAACB to obtain a convolution block RDAACN model.

[0036] In this embodiment, the decoder includes a 3×3 convolution layer with 512 convolution kernels and a rectified linear unit ReLU activation function, a first upsampling module, a 3×3 convolution layer with 64 convolution kernels and a rectified linear unit ReLU activation function, a second upsampling module, a 3×3 convolution layer with 32 convolution kernels and a rectified linear unit ReLU activation function, and a 1×1 convolution layer with 2 convolution kernels.

[0037] In this embodiment, the method in which the decoder outputs the prediction result according to the image features extracted by the encoder is: Step S211: After the feature map output by the convolution block RDAACB4 is input into the 3×3 convolution layer of the decoder whose number of convolution kernels is 512 and whose activation function is the rectified linear unit ReLU, the output feature map of the 3×3 convolution layer is upsampled by 2 times, and then input into the 3×3 convolution layer whose number of convolution kernels is 64 and whose activation function is the rectified linear unit ReLU. The obtained output feature map is added to the output feature map of the convolution block RDAACB2 to obtain a feature fusion map, and then upsampled by 4 times; Step S212: The feature fusion image after 4 times upsampling is input into the 3×3 convolution layer with 32 convolution kernels and the activation function of the rectified linear unit ReLU, and then input into the 1×1 convolution layer with 2 convolution kernels to obtain the output prediction of the RDAACN deep learning model.

[0038] like Figure 4 As shown, in a specific application embodiment, the method for the decoder of the RDAACN deep learning model to output predictions based on the features extracted by the encoder includes the following steps: Step 1: The feature map output by the convolution block RDAACB4 is input into a 3×3 convolution layer with 512 convolution kernels and ReLU activation function. The output feature map of the 3×3 convolution layer is upsampled by 2 times using bilinear interpolation to expand its size by 2 times. Step 2: The feature map output by the first 2x upsampling is input to a 3×3 convolutional layer with 64 convolution kernels and ReLU activation function. The output feature map of the convolutional layer is added to the output feature map of RDAACB2, and then the added feature map is upsampled 4x using bilinear interpolation to expand its size by 4x. Step 3: The feature map of the second 4x upsampling output is input into a 3×3 convolutional layer with 32 convolution kernels and ReLU activation function. The output of the 3×3 convolutional layer is then input into a 1×1 convolutional layer with 2 convolution kernels to obtain the output prediction of RDAACN.

[0039] Step S03: Use the test set to test the trained RDAACN deep learning model to identify the smoke area in the image.

[0040] The outdoor scene smoke area recognition method based on deep learning in this embodiment has the advantages of simple principle and high intelligence, which can effectively improve the recognition efficiency and accuracy of smoke images and can be applied to the fields of computer vision and deep learning. The present invention innovatively adopts a low-Parameters RDAACN model. The convolution block RDAACB constructed in the RDAACN model has dense hole convolution, a large receptive field, and dense pixel sampling. At the same time, an asymmetric convolution kernel is used to reduce the number of parameters, and RDAACB can retain the information of the input feature map; the decoder of the RDAACN model outputs the extraction result of the smoke area in the image with pixel-level accuracy according to the image features extracted by the convolution block RDAACB, thereby identifying the smoke area in the outdoor scene image.

[0041] This embodiment also provides an outdoor scene smoke area recognition device based on deep learning, including a processor and a memory, the memory is used to store an outdoor scene smoke area recognition program based on deep learning, and the processor is used to execute an outdoor scene smoke area recognition method based on deep learning.

[0042] It is understandable that the above method of this embodiment can be executed by a single device, such as a computer or server, etc., and can also be applied to a distributed scenario and completed by multiple devices in cooperation with each other. In the case of a distributed scenario, one of the multiple devices can only execute one or more steps in the above method of this embodiment, and multiple devices interact to complete the above method. The processor can be implemented in the form of a general-purpose CPU, a microprocessor, an application-specific integrated circuit, or one or more integrated circuits, etc., for executing related programs to implement the above method of this embodiment. The memory can be implemented in the form of a read-only memory ROM, a random access memory RAM, a static storage device, and a dynamic storage device. The memory can store an operating system and other applications. When the above method of this embodiment is implemented by software or firmware, the relevant program code is stored in the memory and called and executed by the processor.

[0043] This embodiment also provides a computer-readable storage medium storing a deep learning-based outdoor scene smoke area recognition program, and the computer program / instructions are programmed or configured to execute the deep learning-based outdoor scene smoke area recognition method through a processor.

[0044] Those skilled in the art should understand that the above-mentioned embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The present invention is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the functions in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer-readable memory produce a product including an instruction device, which implements the functions specified in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide for implementing the process in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The above are only preferred embodiments of the present invention and are not intended to limit the present invention in any form. Although the present invention has been disclosed as above with preferred embodiments, it is not intended to limit the present invention. Therefore, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the scope of protection of the technical solution of the present invention.

[0045] The above is only a preferred embodiment of the present invention, and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment, it is not intended to limit the present invention. Therefore, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the scope of protection of the technical solution of the present invention.

Claims

1. A method for identifying smoke areas in outdoor scenes based on deep learning, characterized in that: The following steps are involved: Step S01: Acquisition of outdoor scene dataset: Obtain an image dataset containing images of different outdoor scenes and construct a training set and a test set; Step S02: Model construction and training: Constructing an RDAACN deep learning model, the RDAACN deep learning model includes an encoder and a decoder; the encoder includes a plurality of sequentially connected convolution blocks RDAACB, the convolution block RDAACB includes a plurality of asymmetric convolution layers with different dilation rates, for extracting image features; the decoder is used to output a prediction result according to the image features extracted by the encoder, the asymmetric convolution layer means that the convolution kernel has different lengths and widths; Using the training set and the test set to train the constructed RDAACN deep learning model, and saving the trained RDAACN deep learning model; Step S03: inputting the outdoor scene image to be identified into the trained RDAACN deep learning model to identify the smoke area in the outdoor scene image to be identified; The plurality of sequentially connected convolution blocks RDAACB in the encoder include sequentially connected convolution blocks RDAACB1, convolution blocks RDAACB2, convolution blocks RDAACB3 and convolution blocks RDAACB4; The parameter calculation formula of the convolution block RDAACB4 is: ， Among them, Parameters represents the parameter quantity, F CH represents the height of the convolution kernel, F CW represents the width of the convolution kernel, C I Indicates the number of channels of the input feature map, C O Indicates the number of channels of the output feature map.

2. The method for outdoor scene smoke area recognition based on deep learning according to claim 1 is characterized in that: In step S02, the convolution block RDAACB includes a first branch for generating a first feature map and a second branch for generating a second feature map, and the first feature map and the second feature map are added to obtain an output feature map.

3. The outdoor scene smoke area recognition method based on deep learning according to claim 2 is characterized in that: The method for obtaining the output feature map comprises: Obtain the original feature map or input image, and input the original feature map or input image into the first branch. The number of convolution kernels is N e / 2, the expansion rate is 3, the output is processed by batch normalization, and the activation function is the 1×3 hole convolution layer with the rectified linear unit ReLU; the output feature map of the 1×3 hole convolution layer is input to the number of convolution kernels N e / 2, expansion rate is 3, output uses batch normalization, activation function uses rectified linear unit ReLU 3×1 hole convolution layer; the output feature map of 3×1 hole convolution layer has the number of input convolution kernels is N e / 2, the expansion rate is 6, the output is processed using batch normalization, and the activation function is a 3×3 hollow convolution layer with a rectified linear unit ReLU; The output feature maps of the 3×1 dilated convolution layer and the 3×3 dilated convolution layer are fused by channel merging, and the fused feature maps are input into the convolution kernel with the number of N e , the activation function uses a 1×1 convolution layer of Sigmoid, and the output feature map of the 1×1 convolution layer is multiplied by the fused feature map to obtain a first feature map; The original feature map or input image is input into the second branch. The number of convolution kernels is N e , the output is processed by batch normalization, the activation function is a 1×3 convolution layer with a rectified linear unit ReLU, and the output feature map of the 1×3 convolution layer is then input to the convolution kernel with the number of N e , the output is processed by batch normalization, and the activation function is a 3×1 convolution layer with a rectified linear unit ReLU to obtain the second feature map; The first feature map is added to the second feature map to obtain an output feature map of the convolution block RDAACB.

4. The method for outdoor scene smoke area recognition based on deep learning according to claim 1, characterized in that: The output ends of the convolution blocks RDAACB1, RDAACB2 and RDAACB3 are also connected to a maximum pooling layer with a window size of 2×2 and a pooling window step of 2 for outputting the output feature map after maximum pooling. The output end of the convolution block RDAACB4 does not include a maximum pooling layer but directly outputs the output feature map.

5. The method for outdoor scene smoke area recognition based on deep learning according to claim 4 is characterized in that: The calculation formula of the parameter amount of the convolution block RDAACB1 to the convolution block RDAACB3 is: ， Among them, Parameters represents the parameter quantity, F CH represents the height of the convolution kernel, F CW represents the width of the convolution kernel, C I Indicates the number of channels of the input feature map, C O Indicates the number of channels of the output feature map.

6. The method for outdoor scene smoke area recognition based on deep learning according to claim 1, characterized in that: In the step S02, the decoder includes a 3×3 convolution layer with 512 convolution kernels and an activation function of a rectified linear unit ReLU, a first upsampling module, a 3×3 convolution layer with 64 convolution kernels and an activation function of a rectified linear unit ReLU, a second upsampling module, a 3×3 convolution layer with 32 convolution kernels and an activation function of a rectified linear unit ReLU, and a 1×1 convolution layer with 2 convolution kernels.

7. The method for outdoor scene smoke area recognition based on deep learning according to claim 6, characterized in that: The method in which the decoder outputs a prediction result according to the image features extracted by the encoder is: Step S211: After the feature map output by the convolution block RDAACB4 is input into the 3×3 convolution layer of the decoder, whose number of convolution kernels is 512 and whose activation function is the rectified linear unit ReLU, the output feature map of the 3×3 convolution layer is upsampled by 2 times, and then input into the 3×3 convolution layer, whose number of convolution kernels is 64 and whose activation function is the rectified linear unit ReLU. The obtained output feature map is added to the output feature map of the convolution block RDAACB2, and the feature fusion map is obtained, and then upsampled by 4 times; Step S212: The feature fusion image after 4 times upsampling is input into a 3×3 convolution layer with 32 convolution kernels and ReLU activation function, and then input into a 1×1 convolution layer with 2 convolution kernels to obtain the output prediction of the RDAACN deep learning model.

8. The method for outdoor scene smoke area recognition based on deep learning according to any one of claims 1 to 7, characterized in that: The number of parameters of the RDAACN deep learning model is the sum of the parameters of all convolutional blocks.

Images