Road recognition and extraction method based on improved DeepLabv3 + unmanned aerial vehicle inspection scene

Abstract

The invention provides a road recognition and extraction method based on an improved DeepLabv3 + unmanned aerial vehicle inspection scene, which improves an existing cascaded cavity space pyramid pooling module and is provided with a feature extraction module, an up-sampling module and a plurality of convolution modules. The input data after feature extraction is processed by a cascade cavity space pyramid pooling module and then is subjected to up-sampling processing, the data after up-sampling processing and the input data after feature extraction are subjected to convolution processing and up-sampling processing again, and a final result is determined. The final result can clean and distinguish the content belonging to the road surface part and not belonging to the road surface part in the picture of the input data, and necessary information support is provided for subsequent unmanned aerial vehicle control and processing.

Description

Technical Field

[0001] The present application relates to the field of drone inspection technology, and in particular to a road recognition and extraction method in drone inspection scenarios based on improved DeepLabv3+. Background Art

[0002] With the rapid development of my country's economy, the construction of highways in my country is very rapid, and the safety hazards of highways are also increasing. Therefore, it is necessary to regularly inspect highways and check for safety hazards. At present, highway inspections are basically carried out manually, with low processing efficiency and high costs, which is difficult to meet the current highway inspection efficiency requirements. In particular, highways often cross mountainous areas with complex terrain, which greatly increases the difficulty of manual inspections.

[0003] Therefore, how to efficiently carry out daily inspections on highways has become an urgent problem to be solved. Summary of the invention

[0004] The present invention provides a road recognition and extraction method based on the improved DeepLabv3+ UAV inspection scenario, which solves the problem of requiring a large amount of computing resources and a large number of skilled samples during the UAV highway inspection process. This method reduces the demand for training samples, reduces the consumption of computing resources, and improves the model training speed. The specific contents of this solution are as follows:

[0005] A road recognition and extraction method in a drone inspection scenario based on an improved DeepLabv3+, characterized in that the method comprises:

[0006] The drone image dataset is input into the feature extraction network for feature extraction to obtain the drone image feature dataset;

[0007] Inputting the drone image feature data set into a cascaded atrous spatial pyramid pooling module to obtain a cascaded data set, wherein the cascaded atrous spatial pyramid pooling module is used to perform multi-layer convolution processing and pooling processing on the drone image feature data set;

[0008] Performing upsampling processing on the cascaded data set to obtain a cascaded upsampled data set;

[0009] Performing convolution processing on the drone image feature dataset to obtain a convolution feature dataset;

[0010] The cascade up-sampled data set and the first convolution feature data set are convolved, and the convolution result is up-sampled to determine the road surface part and the non-road surface part of the image taken by the drone in the drone image data set.

[0011] In another embodiment of the present invention, the inputting the drone image feature dataset into a cascaded hole space pyramid pooling module to obtain a cascaded dataset comprises:

[0012] Performing a first convolution process on the drone image feature data set to obtain a first convolution result;

[0013] Performing a secondary convolution process based on the drone image feature dataset and the first convolution result to obtain a second convolution result;

[0014] By analogy, after N convolutions, the Nth convolution result is obtained, where N ≥ 2;

[0015] Performing pooling processing on the Nth convolution result and the drone image feature dataset to obtain a pooling result;

[0016] Convolution processing is performed on the first convolution result to the Nth convolution result and the pooling result to obtain the cascade data set.

[0017] In another embodiment of the present invention, the feature extraction layer network includes: an image input layer, a feature convolution layer, a bottleneck layer, a global average pooling, a fully connected layer, and an output layer; the image input layer, the feature convolution layer, the bottleneck layer, the global average pooling, the fully connected layer, and the output layer are connected in sequence.

[0018] The drone image feature dataset is subjected to feature extraction processing in the feature convolution layer to obtain an initial feature dataset;

[0019] The initial feature data set is processed by point-by-point convolution, depth-wise convolution, and full-connection mapping in the bottleneck layer to obtain a bottleneck feature data set;

[0020] The bottleneck feature data set is pooled in the global average pooling to obtain a pooled feature data set;

[0021] After the pooled feature dataset is processed by the fully connected layer, the drone image feature dataset is obtained.

[0022] In another embodiment of the present invention, the bottleneck layer includes: a first 1X1 point-by-point convolution layer, a 3X3 depth convolution layer, a bottleneck pooling layer, a bottleneck fully connected layer, and a second 1X1 point-by-point convolution layer; the first 1X1 point-by-point convolution layer, the 3X3 depth convolution layer, the bottleneck pooling layer, the bottleneck fully connected layer, and the second 1X1 point-by-point convolution layer are connected in sequence;

[0023] The initial feature data set is convolved in the first 1×1 point-by-point convolution layer and the 3×3 depth convolution layer to obtain an initial convolution data set;

[0024] The initial convolutional data set is pooled in the bottleneck pooling layer to obtain an initial pooled data set;

[0025] The initial pooled data set is weightedly mapped to the second 1X1 point-by-point convolutional layer in the bottleneck fully connected layer, and convolution processing is performed in the second 1X1 point-by-point convolutional layer to obtain the bottleneck feature data set.

[0026] In another embodiment of the present invention, before inputting the drone image dataset into the feature extraction network to extract features and obtain the drone image feature dataset, the method further includes:

[0027] Use a flying drone to shoot the inspection scene road, obtain the drone inspection scene road image and perform calibration;

[0028] The drone inspection scene road image is subjected to data enhancement processing and image expansion to determine the drone image data set.

[0029] In another embodiment of the present invention, the step of acquiring and calibrating a road image of a drone inspection scene includes:

[0030] The road image of the drone inspection scene is binarized, the road is marked as a class and set to a first grayscale value; objects outside the road are marked as a background class and set to a second grayscale value.

[0031] In another embodiment of the present invention, performing data enhancement processing and image expansion on the drone inspection scene road image to determine the drone image data set includes:

[0032] The road image of the drone inspection scene is enhanced by performing one or more of the four operations of translation, mirroring, scaling and rotation.

[0033] In another embodiment of the present invention, the translation, mirroring, scaling, and rotation operations are specifically:

[0034] The translation operation is: moving the drone inspection road image and label vertically or / and horizontally by a number of pixels;

[0035] The mirror operation is: flipping the drone inspection road image and label along the longitudinal center line of the image, or flipping along the transverse center line of the image;

[0036] The zoom operation is: randomly zooming in or out the drone inspection road image and label by a certain multiple;

[0037] The rotation operation is: randomly rotating the drone inspection road image and label by a certain angle clockwise.

[0038] In another embodiment of the present invention, five convolutional layers and one pooling layer are arranged in the cascaded atrous spatial pyramid pooling module, namely, the first convolutional layer, the second convolutional layer, the third convolutional layer, the fourth convolutional layer, and the fifth convolutional layer; wherein, the first convolutional layer is a 1X1 convolutional layer; the second convolutional layer is a 3X3 convolutional layer, and the hole rate is: 6; the third convolutional layer is a 3X3 convolutional layer, and the hole rate is: 12; the fourth convolutional layer is a 3X3 convolutional layer, and the hole rate is: 18; the fifth convolutional layer is a 3X3 convolutional layer, and the hole rate is: 24.

[0039] In another embodiment of the present invention, 80% of the drone image data set is a training set, and 20% is a validation set.

[0040] The beneficial effects of the present invention are as follows:

[0041] The present invention provides a road recognition and extraction method in an unmanned aerial vehicle inspection scene based on an improved DeepLabv3+, improves the existing cascaded void space pyramid pooling module, and is provided with a feature extraction module, an upsampling module, and multiple convolution modules, so that the input data after feature extraction is processed by the cascaded void space pyramid pooling module and then subjected to upsampling processing, and the upsampling data is convoluted and upsampled again with the input data after feature extraction to determine the final result. The final result can clean and distinguish the content belonging to the road surface part and the content not belonging to the road surface part in the picture of the input data, and provide necessary information support for subsequent unmanned aerial vehicle control and processing.

[0042] The neural network model in the feature extraction module has a smaller size than the existing neural network model, and can use fewer training samples and input data to obtain effective feature extraction results. Compared with the existing technology, the improved model reduces the training time by 15%, the average intersection-over-union ratio is increased from 82.3% of the existing technology to 84.1%, and the average pixel accuracy of the category is increased from 88.96% to 90.20%. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is an example image after the road image is binarized in the present invention;

[0044] Figure 2 is an example diagram of data enhancement after the road image is binarized in the present invention;

[0045] Figure 3 It is the Xception network structure diagram involved in the present invention;

[0046] Figure 4 It is a schematic diagram of the feature recognition network structure in the present invention;

[0047] Figure 5It is a structural diagram of the cascaded hollow space pyramid pooling module designed by the present invention;

[0048] Figure 6 This is the improved DeepLabv3+ network structure diagram in the present invention. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solutions and advantages of this specification clearer, the technical solutions of this application will be clearly and completely described below in conjunction with the specific embodiments of this application and their corresponding drawings. Obviously, the described embodiments are only part of the embodiments of this specification, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this document.

[0050] The following combination Figure 1-6 , describes in detail the technical solutions provided by each embodiment of this specification. Specific embodiment 1:

[0052] The present invention provides a road recognition and extraction method in an unmanned aerial vehicle inspection scenario based on an improved DeepLabv3+. The method is applied to highway inspection unmanned aerial vehicles, and is specifically used to determine the part of the highway image that belongs to the road surface and the part that does not belong to the road surface, thereby providing necessary information support for subsequent unmanned aerial vehicle control and operation. In the existing unmanned aerial vehicle highway inspection method, a complete convolutional neural network model is adopted to process image information by embedding and calling. However, due to the large size of the existing convolutional neural network model, a large amount of data and computing resources will be occupied, resulting in a long training time and high computing resource consumption in the unmanned aerial vehicle inspection scenario. It will also lead to problems such as insufficient road feature extraction and image segmentation leakage. The method of this scheme modifies and optimizes the existing convolutional neural network model, overcoming the above-mentioned defects of the prior art.

[0053] The specific contents are as follows:

[0054] Step 1: Conduct UAV flight tests, collect 3,000 images of roads in UAV inspection scenes, and calibrate them. The specific method of calibration is: binarize the image, mark the road as one category, set the grayscale value to 90, and classify objects such as buildings, vegetation, and vehicles as background categories, and set the grayscale value to 0. An example of a road image after binarization is shown below: Figure 1 shown.

[0055] Step 2: Data enhancement processing, expand the 3,000 drone images into a data set of 18,000 images to complete the data set production. In deep learning, a small sample size of the data set will cause overfitting of the network training, that is, it fits well to the current training data set, but performs poorly on the validation set and test set. In order to avoid overfitting, methods such as translation, mirroring, scaling, rotation, and random combination and superposition of four operations are used to enhance the data and expand the number of data set samples. Specifically, the translation operation is to translate the image and label up and down by a number of pixels (a random number less than half the image height), and to the left and right by a number of pixels (a random number less than half the image width); the mirroring operation is divided into flipping in the horizontal or vertical direction, and the horizontal flipping is to flip the image and label along the longitudinal center line of the image, and the vertical flipping is to flip the image and label along the horizontal center line of the image; the scaling operation randomly enlarges or reduces the image by several times (1 to 5 times); the rotation operation randomly rotates the image and label clockwise by several angles (0° to 360°). An example of a road image after data enhancement is shown below. Figure 2 shown.

[0056] Through data enhancement processing, on the one hand, the overfitting phenomenon caused by the small sample size of the data set in the neural network is overcome; on the other hand, the amount of image capture by the drone is reduced, which speeds up the training of the model.

[0057] Step 3: Training and testing the improved DeepLabv3+ deep neural network. The present invention mainly improves and optimizes the network structure of DeepLabv3+, and the specific optimization contents are as follows:

[0058] Optimization 1: Improve network structure to increase model training speed

[0059] Since the Xception network structure used in the original DeepLabv3+ algorithm is very complex and has a large number of parameters, it consumes a lot of training time and computing resources, making it unsuitable for drone inspection scenarios that require high real-time performance. In order to improve model training efficiency and reduce computing resource consumption, this solution optimizes the Xception network structure, reduces the number of network parameters by using deep separable convolution, and improves the speed of road image feature extraction and model performance.

[0060] The structure of the Xception network is based on the Inception network with the addition of deep separable convolutions, and is improved with the residual network structure. The deep separable convolutions used in the network completely decouple the spatial correlation and cross-channel correlation in the feature tensor, which can give full play to the role of network parameters compared to the Inception network. Figure 3As shown in the figure, there are 36 convolutional layers, including 1 downsampling convolutional layer, 1 regular convolutional layer and 34 depth-separable convolutional layers; divided into 14 modules, the maximum number of channels reaches 2048, which makes the Xception network structure deeper, the parameter calculation is large, the model occupies a large amount of memory, and is not suitable for deployment in mobile terminals and embedded devices for real-time detection. At the same time, the Xception network lacks the fusion of multi-scale features, and may cause feature loss during feature extraction due to the single receptive field.

[0061] The feature extraction network in this solution optimizes the problem of large computational complexity of traditional convolutional neural networks, and adopts lightweight structures such as deep separable convolution and residual modules, so that the model can achieve higher accuracy with lower parameter and computational complexity, and can be deployed in environments with insufficient hardware conditions such as mobile devices and embedded devices. The feature extraction network structure of this solution is as follows: Figure 4 shown.

[0062] There are 3X3 convolutional layers, 15 bottleneck layers, global average pooling, and fully connected layers. The above settings are connected in sequence. The input image enters the feature extraction network from the 3X3 convolutional layer and is output from the fully connected layer after processing.

[0063] The new nonlinear h-swish is introduced in the feature extraction network to replace the original nonlinear swish algorithm, which improves the quality and speed of calculation. The head of the feature extraction network is optimized as follows. In order to reduce latency while retaining high-dimensional features, the average pooling layer is removed and the feature map is calculated using 1x1 convolution. After the feature generation layer is removed, the bottleneck mapping layer of the original Xception network structure is no longer needed, which will reduce the consumption of more computing resources. In addition, the feature extraction network adds an attention mechanism based on the original Xception network structure. The attention mechanism is equivalent to a vector weight assignment mechanism. The main operation is to perform pooling on each channel of the output feature matrix. The number of vectors obtained is the number of channels, and then the output vector is obtained after passing through two circle connection layers. The vector here is equivalent to a weight relationship analyzed for each input channel.

[0064] Through the above simplification and optimization, the structure of the existing Xception network has been greatly reduced, making the model smaller, requiring less computing resources, and faster. However, the computing effect has not decreased but has been improved to a certain extent. See below for specific data.

[0065] Optimization 2: Using cascaded atrous spatial pyramid pooling modules to improve road feature extraction capabilities

[0066] The Atrous Spatial Pyramid Pooling (ASPP) module in the DeepLabv3+ algorithm uses atrous convolution to obtain a larger receptive field, but for road feature extraction tasks, feature loss is prone to occur, resulting in problems such as missed segmentation. In order to improve the model's perception of road targets, the cascaded atrous spatial pyramid pooling (C-ASPP) module is adopted. The C-ASPP module can further extract multi-scale feature information of the image by adding a convolutional layer; then by cascading different branches, the feature output of the previous layer is fused with the feature input of the next layer as the feature input of the next layer, to fuse multi-level cross-scale features to obtain more comprehensive semantic information, improve the utilization of pixel information, and thus strengthen the extraction of global information and detailed features of the road, such as Figure 5 .

[0067] The improved deep neural network structure is as follows Figure 6 . 80% of the drone image dataset was divided into a training set and 20% into a validation set for testing. The experimental results show that under the same number of iterations, compared with the DeepLabv3 network, the improved DeepLabv3+ network training time is reduced by 15%, mIOU (mean intersection over union) is increased from 82.3% to 84.1%, and MPA (mean pixel accuracy) is increased from 88.96% to 90.20%. Specific embodiment 2:

[0069] The present invention provides a road recognition and extraction method based on the improved DeepLabv3+ UAV inspection scenario, which solves the problem of requiring a large amount of computing resources and a large number of skilled samples during the UAV highway inspection process. This method reduces the demand for training samples, reduces the consumption of computing resources, and improves the model training speed. The specific contents of this solution are as follows:

[0070] A road recognition and extraction method in a drone inspection scenario based on an improved DeepLabv3+, characterized in that the method comprises:

[0071] The drone image dataset is input into the feature extraction network for feature extraction to obtain the drone image feature dataset;

[0072] Inputting the drone image feature data set into a cascaded atrous spatial pyramid pooling module to obtain a cascaded data set, wherein the cascaded atrous spatial pyramid pooling module is used to perform multi-layer convolution processing and pooling processing on the drone image feature data set;

[0073] Performing upsampling processing on the cascaded data set to obtain a cascaded upsampled data set;

[0074] Performing convolution processing on the drone image feature dataset to obtain a convolution feature dataset;

[0075] The cascade up-sampled data set and the first convolution feature data set are convolved, and the convolution result is up-sampled to determine the road surface part and the non-road surface part of the image taken by the drone in the drone image data set.

[0076] In another embodiment of the present invention, the inputting the drone image feature dataset into a cascaded hole space pyramid pooling module to obtain a cascaded dataset comprises:

[0077] Performing a first convolution process on the drone image feature data set to obtain a first convolution result;

[0078] Performing a secondary convolution process based on the drone image feature dataset and the first convolution result to obtain a second convolution result;

[0079] By analogy, after N convolutions, the Nth convolution result is obtained, where N ≥ 2;

[0080] Performing pooling processing on the Nth convolution result and the drone image feature dataset to obtain a pooling result;

[0081] Convolution processing is performed on the first convolution result to the Nth convolution result and the pooling result to obtain the cascade data set.

[0082] In another embodiment of the present invention, the feature extraction layer network includes: an image input layer, a feature convolution layer, a bottleneck layer, a global average pooling, a fully connected layer, and an output layer; the image input layer, the feature convolution layer, the bottleneck layer, the global average pooling, the fully connected layer, and the output layer are connected in sequence.

[0083] The drone image feature dataset is subjected to feature extraction processing in the feature convolution layer to obtain an initial feature dataset;

[0084] The initial feature data set is processed by point-by-point convolution, depth-wise convolution, and full-connection mapping in the bottleneck layer to obtain a bottleneck feature data set;

[0085] The bottleneck feature data set is pooled in the global average pooling to obtain a pooled feature data set;

[0086] After the pooled feature dataset is processed by the fully connected layer, the drone image feature dataset is obtained.

[0087] In another embodiment of the present invention, the bottleneck layer includes: a first 1X1 point-by-point convolution layer, a 3X3 depth convolution layer, a bottleneck pooling layer, a bottleneck fully connected layer, and a second 1X1 point-by-point convolution layer; the first 1X1 point-by-point convolution layer, the 3X3 depth convolution layer, the bottleneck pooling layer, the bottleneck fully connected layer, and the second 1X1 point-by-point convolution layer are connected in sequence;

[0088] The initial feature data set is convolved in the first 1×1 point-by-point convolution layer and the 3×3 depth convolution layer to obtain an initial convolution data set;

[0089] The initial convolutional data set is pooled in the bottleneck pooling layer to obtain an initial pooled data set;

[0090] The initial pooled data set is weightedly mapped to the second 1X1 point-by-point convolutional layer in the bottleneck fully connected layer, and convolution processing is performed in the second 1X1 point-by-point convolutional layer to obtain the bottleneck feature data set.

[0091] In another embodiment of the present invention, before inputting the drone image dataset into the feature extraction network to extract features and obtain the drone image feature dataset, the method further includes:

[0092] Use a flying drone to shoot the inspection scene road, obtain the drone inspection scene road image and perform calibration;

[0093] The drone inspection scene road image is subjected to data enhancement processing and image expansion to determine the drone image data set.

[0094] In another embodiment of the present invention, the step of acquiring and calibrating a road image of a drone inspection scene includes:

[0095] The road image of the drone inspection scene is binarized, the road is marked as a class and set to a first grayscale value; objects outside the road are marked as a background class and set to a second grayscale value.

[0096] In another embodiment of the present invention, performing data enhancement processing and image expansion on the drone inspection scene road image to determine the drone image data set includes:

[0097] The road image of the drone inspection scene is enhanced by performing one or more of the four operations of translation, mirroring, scaling and rotation.

[0098] In another embodiment of the present invention, the translation, mirroring, scaling, and rotation operations are specifically:

[0099] The translation operation is: moving the drone inspection road image and label vertically or / and horizontally by a number of pixels;

[0100] The mirror operation is: flipping the drone inspection road image and label along the longitudinal center line of the image, or flipping along the transverse center line of the image;

[0101] The zoom operation is: randomly zooming in or out the drone inspection road image and label by a certain multiple;

[0102] The rotation operation is: randomly rotating the drone inspection road image and label by a certain angle clockwise.

[0103] In another embodiment of the present invention, five convolutional layers and one pooling layer are arranged in the cascaded atrous spatial pyramid pooling module, namely, the first convolutional layer, the second convolutional layer, the third convolutional layer, the fourth convolutional layer, and the fifth convolutional layer; wherein, the first convolutional layer is a 1X1 convolutional layer; the second convolutional layer is a 3X3 convolutional layer, and the hole rate is: 6; the third convolutional layer is a 3X3 convolutional layer, and the hole rate is: 12; the fourth convolutional layer is a 3X3 convolutional layer, and the hole rate is: 18; the fifth convolutional layer is a 3X3 convolutional layer, and the hole rate is: 24.

[0104] In another embodiment of the present invention, 80% of the drone image data set is a training set, and 20% is a validation set.

[0105] This paper proposes a road recognition and extraction method based on the improved DeepLabv3+ UAV inspection scenario. The improved network adopts an encoder-decoder structure. The encoder part uses a lightweight feature extraction network structure as the backbone feature extraction network, which improves the ability to extract highway road features while improving the model training speed; by adopting the Cascade Atrous Spatial Pyramid Pooling (C-ASPP) module, the model's perception of road targets and the ability to extract road detail features are improved. The road recognition and extraction method based on the improved DeepLabv3+ UAV inspection scenario solves the problems of long training time, high computing resource consumption, and insufficient road feature extraction leading to missed segmentation in the original DeepLabv3+ algorithm in the inspection scenario.

Claims

1. A road recognition and extraction method based on improved DeepLabv3+ UAV inspection scenario, characterized in that: The method comprises: The drone image dataset is input into the feature extraction network for feature extraction to obtain the drone image feature dataset; Inputting the drone image feature data set into a cascaded atrous spatial pyramid pooling module to obtain a cascaded data set, wherein the cascaded atrous spatial pyramid pooling module is used to perform multi-layer convolution processing and pooling processing on the drone image feature data set; Performing upsampling processing on the cascaded data set to obtain a cascaded upsampled data set; Performing convolution processing on the UAV image feature data set to obtain a convolution feature data set; The cascade up-sampled data set and the first convolution feature data set are convolved, and the convolution result is up-sampled to determine the road surface part and the non-road surface part of the image taken by the drone in the drone image data set.

2. The method for road recognition and extraction in a drone inspection scenario based on an improved DeepLabv3+ as claimed in claim 1, characterized in that: The drone image feature dataset is input into the cascaded hole space pyramid pooling module to obtain a cascaded dataset including: Performing a first convolution process on the drone image feature data set to obtain a first convolution result; Performing a secondary convolution process based on the drone image feature dataset and the first convolution result to obtain a second convolution result; By analogy, after N convolutions, the Nth convolution result is obtained, where N ≥ 2; Performing pooling processing on the Nth convolution result and the drone image feature data set to obtain a pooling result; Convolution processing is performed on the first convolution result to the Nth convolution result and the pooling result to obtain the cascade data set.

3. The road recognition and extraction method based on the improved DeepLabv3+ UAV inspection scenario as claimed in claim 1 is characterized in that: The feature extraction layer network includes: an image input layer, a feature convolution layer, a bottleneck layer, a global average pooling, a fully connected layer, and an output layer; the image input layer, the feature convolution layer, the bottleneck layer, the global average pooling, the fully connected layer, and the output layer are connected in sequence. The drone image feature dataset is subjected to feature extraction processing in the feature convolution layer to obtain an initial feature dataset; The initial feature data set is processed by point-by-point convolution, depth-wise convolution, and full-connection mapping in the bottleneck layer to obtain a bottleneck feature data set; The bottleneck feature data set is pooled in the global average pooling to obtain a pooled feature data set; After the pooled feature dataset is processed by the fully connected layer, the drone image feature dataset is obtained.

4. The method for road recognition and extraction in a drone inspection scenario based on an improved DeepLabv3+ as claimed in claim 3 is characterized in that: The bottleneck layer includes: a first 1X1 point-by-point convolution layer, a 3X3 depth convolution layer, a bottleneck pooling layer, a bottleneck fully connected layer, and a second 1X1 point-by-point convolution layer; the first 1X1 point-by-point convolution layer, the 3X3 depth convolution layer, the bottleneck pooling layer, the bottleneck fully connected layer, and the second 1X1 point-by-point convolution layer are connected in sequence; The initial feature data set is convolved in the first 1×1 point-by-point convolution layer and the 3×3 depth convolution layer to obtain an initial convolution data set; The initial convolutional data set is pooled in the bottleneck pooling layer to obtain an initial pooled data set; The initial pooled data set is weightedly mapped to the second 1X1 point-by-point convolutional layer in the bottleneck fully connected layer, and convolution processing is performed in the second 1X1 point-by-point convolutional layer to obtain the bottleneck feature data set.

5. The method for road recognition and extraction in a drone inspection scenario based on an improved DeepLabv3+ as claimed in claim 1, characterized in that: Before inputting the drone image dataset into the feature extraction network to extract features and obtain the drone image feature dataset, the following step is also included: Use a flying drone to shoot the inspection scene road, obtain the drone inspection scene road image and perform calibration; The drone inspection scene road image is subjected to data enhancement processing and image expansion to determine the drone image data set.

6. The method for road recognition and extraction in a drone inspection scenario based on an improved DeepLabv3+ as claimed in claim 5, characterized in that: The method of obtaining and calibrating the road image of the drone inspection scene includes: The road image of the drone inspection scene is binarized, the road is marked as a class and set to a first grayscale value; objects outside the road are marked as a background class and set to a second grayscale value.

7. The method for road recognition and extraction in a drone inspection scenario based on an improved DeepLabv3+ as claimed in claim 5, characterized in that: The performing data enhancement processing and image expansion on the drone inspection scene road image to determine the drone image data set includes: The road image of the drone inspection scene is enhanced by superimposing one or more of the four operations of translation, mirroring, scaling and rotation to enhance data.

8. The road recognition and extraction method based on the improved DeepLabv3+ UAV inspection scenario as claimed in claim 7 is characterized in that: The specific operations of translation, mirroring, scaling, and rotation are as follows: The translation operation is: moving the drone inspection road image and label vertically or / and horizontally by a number of pixels; The mirror operation is: flipping the drone inspection road image and label along the longitudinal center line of the image, or flipping along the transverse center line of the image; The zoom operation is: randomly zooming in or out the drone inspection road image and label by a certain multiple; The rotation operation is: randomly rotating the drone inspection road image and label by a certain angle clockwise.

9. The method for road recognition and extraction in a drone inspection scenario based on an improved DeepLabv3+ as claimed in claim 1, characterized in that: The cascaded atrous spatial pyramid pooling module is provided with five convolutional layers and one pooling layer, which are respectively the first convolutional layer, the second convolutional layer, the third convolutional layer, the fourth convolutional layer, and the fifth convolutional layer; wherein the first convolutional layer is a 1X1 convolutional layer; the second convolutional layer is a 3X3 convolutional layer, and the atrous rate is: 6; the third convolutional layer is a 3X3 convolutional layer, and the atrous rate is: 12; the fourth convolutional layer is a 3X3 convolutional layer, and the atrous rate is: 18; the fifth convolutional layer is a 3X3 convolutional layer, and the atrous rate is:

24.

10. The method for road recognition and extraction in a drone inspection scenario based on an improved DeepLabv3+ as claimed in claim 1, characterized in that: 80% of the drone image dataset is a training set and 20% is a validation set.

Images