A method, system, device and storage medium for locating a navigation data table
By using grouping convolution, reverse bottleneck structure, SE attention mechanism and feature extraction of feature pyramid structure in navigation data table positioning, combined with the cubic convolution interpolation algorithm, the problem of insufficient feature extraction in the existing technology is solved, and the The accuracy of table positioning.
Patent Information
- Application Number
- CN202410147295.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-02-01
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2044-02-01
AI Technical Summary
In the table positioning of navigation data data, the feature extraction layer adopts the ResNet model, resulting in insufficient feature details and a single-layer feature map. After ROIPooling processing, it is difficult to accurately locate the table position.
Feature extraction is performed based on grouping convolution, reverse bottleneck structure, SE attention mechanism and feature pyramid structure, combined with the cubic convolution interpolation algorithm, and abandon the traditional ROI pooling layer and perform feature mapping to improve feature extraction accuracy.
Through multi-size feature extraction and smooth edge processing, the accuracy of navigation data table positioning is improved, feature pixel loss is avoided, and more full feature utilization is achieved.
Smart Images

Figure CN118038476B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of computer technology, and in particular to a method, system, device and storage medium for locating a navigation data table. Background Art
[0002] As the number of passengers choosing to travel by plane increases, the number of flights at each airport has increased significantly, and flight data has become increasingly abundant. Among them, flight data is mainly presented in the form of tables.
[0003] At present, the table positioning technology of navigation data mainly adopts the Faster R-CNN (Faster Region-Convolutional Neural Network) algorithm, but the original feature extraction layer of the algorithm adopts the ResNet (Residual Neural Network) model. The extracted feature details are very few, and they are all single-layer feature maps. Moreover, after being rounded twice by the original ROIPooling (Region of interest pooling), it will cause a loss of accuracy, making it difficult to accurately locate the position of the table in the navigation data. Summary of the invention
[0004] The present invention provides a navigation data table positioning method, system, device and storage medium to solve the problem that the prior art cannot accurately position the position of the table in the navigation data.
[0005] To achieve the above object, an embodiment of the present invention provides a method for locating a navigation data table, comprising:
[0006] Preprocess the navigation data to obtain a navigation data image;
[0007] Based on group convolution, reverse bottleneck structure, SE attention mechanism and feature pyramid structure, feature extraction is performed on the navigation data image to obtain a feature map;
[0008] The feature map is input into the RPN region proposal network to locate the target table and obtain a proposal box;
[0009] According to the feature map and the suggestion box, a cubic convolution interpolation algorithm is used to obtain a suggestion feature map;
[0010] The proposed feature map is subjected to bounding box regression processing to obtain a table area of the navigation data image.
[0011] As an improvement of the above scheme, the feature extraction of the navigation data picture is performed based on the grouped convolution, reverse bottleneck structure, SE attention mechanism and feature pyramid structure to obtain a feature map, including:
[0012] Extracting a first feature and a first feature map from the navigation data image based on grouped convolution and reverse bottleneck structure;
[0013] Combining the SE attention mechanism and the feature pyramid structure, the first feature is upsampled to obtain the second feature, and the first feature map is convolved to increase the number of channels to the same as the number of channels of the convolution kernel to obtain the second feature map, and then the second feature and the second feature map are directly added;
[0014] After repeating the above steps several times, several feature maps of different scales are obtained.
[0015] As an improvement of the above scheme, the formula of the SE attention mechanism is:
[0016] out=σ(W2δ(W1in))
[0017] Among them, out represents the obtained feature weights of each channel, in represents the feature map of the input attention; σ() represents the Sigmoid activation function, δ() represents the ReLu activation function, W1 represents the parameters of the first fully connected layer, and W2 represents the parameters of the second fully connected layer.
[0018] As an improvement of the above solution, the feature pyramid structure includes a bottom-up path, a top-down path and a lateral connection.
[0019] As an improvement of the above solution, the feature map is input into the RPN region proposal network to locate the target table to obtain the suggestion box, including:
[0020] Input the feature map into the RPN region proposal network, classify the anchors of the feature map by a normalized exponential function based on the RPN region proposal network, obtain positive anchors with targets and negative anchors without targets, and calculate the bounding box regression offset of the corresponding anchors;
[0021] All suggestions are obtained according to the bounding box regression offsets of the positive anchor and the corresponding anchor, and suggestions that do not meet the preset boundary standards are removed to obtain a suggestion box.
[0022] As an improvement of the above solution, the formula of the cubic convolution interpolation algorithm is:
[0023]
[0024] Among them, g(x) represents the interpolation result of the interpolation function at point x, u represents the unit step function, h represents the sampling interval, x represents the position of the interpolation point, k represents the convolution kernel used for interpolation convolution, and x k represents the position of the corresponding interpolation point in the convolution kernel k, c represents the convolution coefficient, c k Represents the coefficient corresponding to the convolution kernel k.
[0025] As an improvement of the above solution, after obtaining the suggested feature map by using a cubic convolution interpolation algorithm according to the feature map and the suggestion box, the navigation data table positioning method further includes:
[0026] A category determination is performed on the suggested feature map to obtain a category of the suggested feature map to serve as a category of the table area.
[0027] To achieve the above purpose, an embodiment of the present invention further provides a navigation data table positioning system, comprising:
[0028] A data preprocessing module is used to preprocess the navigation data to obtain a navigation data image;
[0029] A feature extraction module, used for extracting features from the navigation data image based on grouped convolution, reverse bottleneck structure, SE attention mechanism and feature pyramid structure to obtain a feature map;
[0030] A suggestion box acquisition module is used to input the feature map into the RPN region proposal network to locate the target table and obtain a suggestion box;
[0031] A suggested feature map acquisition module, used to obtain a suggested feature map by using a cubic convolution interpolation algorithm according to the feature map and the suggested box;
[0032] The table positioning module is used to perform bounding box regression processing on the proposed feature map to obtain a table area of the navigation data image.
[0033] To achieve the above-mentioned purpose, an embodiment of the present invention further provides a navigation data table positioning device, comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor implements the above-mentioned navigation data table positioning method when executing the computer program.
[0034] To achieve the above-mentioned purpose, an embodiment of the present invention further provides a computer-readable storage medium, which includes a stored computer program; wherein, when the computer program is running, it controls the device where the computer-readable storage medium is located to execute the above-mentioned navigation data table positioning method.
[0035] Implementing the embodiments of the present invention has the following beneficial effects:
[0036] The navigation data table positioning method, system, device and storage medium provided by the embodiment of the present invention first pre-processes the navigation data to obtain the navigation data picture; then based on the grouped convolution, reverse bottleneck structure, SE attention mechanism and feature pyramid structure, the navigation data picture is feature extracted to obtain a feature map; the feature map is input into the RPN region proposal network to locate the target table to obtain a proposal frame; then, according to the feature map and the proposal frame, a cubic convolution interpolation algorithm is used to obtain a proposal feature map; finally, the proposal feature map is subjected to bounding box regression processing to obtain the table area of the navigation data picture. The embodiment of the present invention extracts multi-scale features of the navigation data picture for training based on the grouped convolution, reverse bottleneck structure, SE attention mechanism and feature pyramid structure. At the same time, the violent rounding method of the ROI pooling layer in the traditional Faster R-CNN model is abandoned, and the cubic convolution interpolation method is introduced for feature mapping, which can produce very smooth edges and avoid the loss of feature map pixels caused by forcibly rounding off in the feature mapping stage, thereby achieving more full utilization of image features, obtaining more accurate feature extraction frames, and improving the accuracy of navigation data table positioning. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] Figure 1 is a flow chart of a method for locating a navigation data table provided by an embodiment of the present invention;
[0038] Figure 2 is a schematic diagram of the conversion of the navigation data file type provided by an embodiment of the present invention;
[0039] Figure 3 is a flowchart of an improved Faster R-CNN provided by an embodiment of the present invention;
[0040] Figure 4 It is a structural comparison diagram of ResNet and ConvNeXt-XL provided by an embodiment of the present invention;
[0041] Figure 5 is a flow chart of obtaining a feature map provided by an embodiment of the present invention;
[0042] Figure 6 is a schematic diagram of table category determination provided by an embodiment of the present invention;
[0043] Figure 7 is a structural block diagram of a navigation data table positioning system provided by an embodiment of the present invention;
[0044] Figure 8 It is a structural block diagram of a navigation data table positioning device provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0045] 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.
[0046] In the description of the present invention, it should be noted that the step numbers in the text are only for the convenience of explanation of specific embodiments and do not serve to limit the order in which the steps are executed. The terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of this application, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined.
[0047] See also Figure 1 , Figure 1 : is a flow chart of a method for locating a navigation data table provided by an embodiment of the present invention. The method for locating a navigation data table includes steps S1 to S5, which are specifically as follows:
[0048] S1. Preprocess the navigation data to obtain a navigation data image;
[0049] S2. Based on group convolution, reverse bottleneck structure, SE attention mechanism and feature pyramid structure, feature extraction is performed on the navigation data image to obtain a feature map;
[0050] S3, inputting the feature map into the RPN region proposal network to locate the target table and obtain a suggestion box;
[0051] S4, according to the feature map and the suggestion box, using a cubic convolution interpolation algorithm to obtain a suggestion feature map;
[0052] S5. Perform bounding box regression processing on the proposed feature map to obtain a table area of the navigation data image.
[0053] It is understandable that if Figure 2 As shown, the navigation data file (i.e., navigation data data) has various types, including pages based on pure text, mixed text and table, and pure table image, which cannot be directly used as input for predicting the table bounding box. Therefore, the navigation data data needs to be preprocessed, that is, converting all formats of navigation data files (Word or PDF) into images (i.e., navigation data data images) as input data.
[0054] For example, Figure 3 As shown, the navigation data is first preprocessed to obtain a navigation data picture as input; then the ConvNeXt-XL model based on grouped convolution, reverse bottleneck structure, SE attention mechanism and feature pyramid structure is used to extract features from the navigation data picture to obtain a feature map; the feature map is input into the RPN region proposal network to locate the target table to obtain a proposal box; then, in the ROI Conv layer, a cubic convolution interpolation algorithm is used to obtain a proposal feature map based on the feature map and the proposal box; finally, a bounding box regression process is performed on the proposal feature map to obtain a table area of the navigation data picture. The embodiment of the present invention extracts multi-size features of the navigation data image for training based on grouped convolution, reverse bottleneck structure, SE attention mechanism and feature pyramid structure. At the same time, the violent rounding method of the ROI pooling layer in the traditional Faster R-CNN model is abandoned, and the cubic convolution interpolation method is introduced for feature mapping, which can produce very smooth edges and avoid the loss of feature map pixels caused by forced rounding in the feature mapping stage, thereby achieving more full utilization of image features, obtaining a more accurate feature extraction frame, and improving the accuracy of navigation data table positioning.
[0055] In an optional embodiment, the feature extraction of the navigation data image is performed based on the grouped convolution, reverse bottleneck structure, SE attention mechanism and feature pyramid structure to obtain a feature map, including:
[0056] Extracting a first feature and a first feature map from the navigation data image based on grouped convolution and reverse bottleneck structure;
[0057] Combining the SE attention mechanism and the feature pyramid structure, the first feature is upsampled to obtain the second feature, and the first feature map is convolved to increase the number of channels to the same as the number of channels of the convolution kernel to obtain the second feature map, and then the second feature and the second feature map are directly added;
[0058] After repeating the above steps several times, several feature maps of different scales are obtained.
[0059] It is understandable that if Figure 4 As shown in FIG. 1 , the feature extractor used by the traditional Faster R-CNN is a ResNet model that is a simple stack of ReLU (Rectified Linear Unit) and BN (Batch Normalization), and its structure is large at both ends and small in the middle.
[0060] In an embodiment of the present invention, the traditional Faster R-CNN is improved, and a grouped convolution ConvNeXt-XL model of GELU (Gaussian Error Linear Unit, nonlinear activation function) and LN (Layer Normalization) is adopted, and a reverse bottleneck structure with small ends and a large middle is introduced to perform feature extraction, which can avoid information loss caused by compressed dimensions when feature information is converted between feature spaces of different dimensions, and extract feature maps with richer information.
[0061] In an optional embodiment, the formula of the SE attention mechanism is:
[0062] out=σ(W2δ(W1in))
[0063] Among them, out represents the obtained feature weights of each channel, in represents the feature map of the input attention; σ() represents the Sigmoid activation function, δ() represents the ReLu activation function, W1 represents the parameters of the first fully connected layer, and W2 represents the parameters of the second fully connected layer.
[0064] For example, Figure 5As shown, the navigation data is converted into a fixed 224×224 size picture as input, and the ConvNeXt-XL feature extraction model is used. The ConvNeXt-XL feature extraction model is based on grouped convolution, reverse bottleneck structure, SE attention mechanism and feature pyramid structure to extract features from the navigation data picture to obtain a feature map, specifically: the features output by the last residual block of each stage are activated, and the feature map (i.e., Feature map×0.5) is output proportionally, i.e., Conv2, Conv3, Conv4 and Conv5, and the outputs of these last residual blocks are represented as {C2, C3, C4, C5}. Among them, since Conv1 occupies a lot of memory, it is not included in the pyramid structure for efficiency considerations. At the same time, the SE (Squeeze-and-Excitation) attention mechanism is introduced to screen out table-irrelevant features or features with low correlation with table features, so that the entire feature extraction model focuses on table effective features. A fusion path is established from the second layer of the model to the fifth layer. The first feature extracted from the fifth layer is upsampled twice (i.e., Feature map×2) to obtain the second feature; the first feature map extracted from the fourth layer is convolved by 1×1 to increase the number of channels to the same number of channels as the convolution kernel to obtain the second feature map; the second feature is directly added to the second feature map, and the above operation is repeated many times to obtain four feature maps of different scales from M2 to M5, represented as {M2, M3, M4, M5}. Finally, the M5 feature map is downsampled with a step size of 2 (i.e., Feature map / 0.5) to generate M6. Using 3×3 convolution on each merged map to generate the final feature map can reduce the confounding effect of upsampling, and the final feature map is called {P2, P3, P4, P5}, corresponding to {C2, C3, C4, C5}, which have the same feature size. In an embodiment of the present invention, after the feature is extracted, it is not directly input into the RPN region proposal network, but first passes through SE attention, which is divided into two stages of compression and excitation. Important feature regions are obtained through channel changes, and the burden brought by the SE attention model is almost negligible.
[0065] In an optional embodiment, the feature pyramid structure includes a bottom-up path, a top-down path and a lateral connection.
[0066] It can be understood that the bottom-up path is a feedforward calculation of the backbone network, which is used to calculate that a feature level is composed of several proportional feature maps. In the embodiment of the present invention, the navigation data table positioning method introduces a feature pyramid structure, which can extract multi-region features, reduce the limitation of the size of the table in the image on the table feature extraction effect, effectively extract multi-scale context information, provide different spatial dimension information for table detection, increase the richness of table image features to a certain extent, and improve the reliability of table feature extraction.
[0067] In an optional embodiment, inputting the feature map into an RPN region proposal network to locate a target table to obtain a suggestion box includes:
[0068] Input the feature map into the RPN region proposal network, classify the anchors of the feature map by a normalized exponential function based on the RPN region proposal network, obtain positive anchors with targets and negative anchors without targets, and calculate the bounding box regression offset of the corresponding anchors;
[0069] All suggestions are obtained according to the bounding box regression offsets of the positive anchor and the corresponding anchor, and suggestions that do not meet the preset boundary standards are removed to obtain a suggestion box.
[0070] It can be understood that the RPN region proposal network is actually divided into two lines for operation. One path is to classify the anchors of the feature map through a normalized exponential function to obtain positive anchors with targets and negative anchors without targets. The anchor is actually a set of matrices, which are used to generate a series of boxes on the pixels of the feature map. Each row has four values (x1, y1, x2, y2), representing the upper left corner coordinates (x1, y1) and the lower right corner coordinates (x2, y2) of the rectangle in the spatial position. The number of anchors varies depending on the size of the input table image. The other path is used to calculate the bounding box regression offset for the anchor to obtain accurate suggestions. The final suggestion layer is responsible for synthesizing the positive anchor and the corresponding bounding box regression offset to obtain all suggestions, while removing suggestions that do not meet the preset boundary standards to obtain suggestion boxes.
[0071] In the embodiment of the present invention, the error of drawing the table bounding box is controlled by fusing the high resolution of low-level features and the high semantic information of high-level features, thereby improving the calculation efficiency and being able to accurately cope with different table features.
[0072] In an optional embodiment, the formula of the cubic convolution interpolation algorithm is:
[0073]
[0074] Among them, g(x) represents the interpolation result of the interpolation function at point x, u represents the unit step function, h represents the sampling interval, x represents the position of the interpolation point, k represents the convolution kernel used for interpolation convolution, and x k represents the position of the corresponding interpolation point in the convolution kernel k, c represents the convolution coefficient, c k Represents the coefficient corresponding to the convolution kernel k.
[0075] It is understandable that the traditional Faster R-CNN uses the ROI pooling layer, and the mapping results are rounded down during feature mapping. Discarding pixels during the rounding process will cause deviations in table detection. As the number of roundings increases, this error will continue to increase.
[0076] In the embodiment of the present invention, the violent rounding method of the ROI pooling layer is abandoned, and the cubic convolution interpolation method is introduced for feature mapping, which can produce very smooth edges and avoid the loss of feature map pixels caused by forcibly rounding off in the feature mapping stage, thereby achieving a more full utilization of image features and obtaining a more accurate feature extraction frame. At the same time, the cubic convolution interpolation is essentially a convolution operation, which is very conducive to the implementation of current hardware, can improve computing efficiency, and save time costs.
[0077] In an optional embodiment, after obtaining the suggested feature map by using a cubic convolution interpolation algorithm according to the feature map and the suggestion box, the navigation data table positioning method further includes:
[0078] A category determination is performed on the suggested feature map to obtain a category of the suggested feature map to serve as a category of the table area.
[0079] It is understandable that if Figure 6 As shown, the table is classified according to the border status, such as a table with a complete border is judged as a wired table, an incomplete border is judged as a table with few lines (including a three-line table, a horizontal line table, etc.), and a table with no border is judged as a wireless table.
[0080] In the embodiment of the present invention, according to the proposed feature map, the specific category of each suggestion is calculated through the fully connected layer and the normalized exponential function, so as to realize the category determination of the proposed feature map. At the same time, the position offset of each suggestion is obtained by bounding box regression again, and finally the position of the table in the navigation data picture is obtained as output.
[0081] The method for locating a navigation data table provided by an embodiment of the present invention first pre-processes the navigation data to obtain a navigation data picture as input; then uses a ConvNeXt-XL model based on grouped convolution, reverse bottleneck structure, SE attention mechanism and feature pyramid structure to extract features from the navigation data picture to obtain a feature map; inputs the feature map into an RPN region proposal network to locate the target table to obtain a proposal box; then, in the ROI Conv layer, a cubic convolution interpolation algorithm is used to obtain a proposal feature map based on the feature map and the proposal box; finally, a bounding box regression process is performed on the proposal feature map to obtain a table area of the navigation data picture. The embodiment of the present invention extracts multi-size features of the navigation data image for training based on grouped convolution, reverse bottleneck structure, SE attention mechanism and feature pyramid structure. At the same time, the violent rounding method of the ROI pooling layer in the traditional Faster R-CNN model is abandoned, and the cubic convolution interpolation method is introduced for feature mapping, which can produce very smooth edges and avoid the loss of feature map pixels caused by forced rounding in the feature mapping stage, thereby achieving more full utilization of image features, obtaining a more accurate feature extraction frame, and improving the accuracy of navigation data table positioning.
[0082] See also Figure 7 , Figure 7 : is a structural block diagram of a navigation data table positioning system 10 provided in an embodiment of the present invention, wherein the navigation data table positioning system 10 comprises:
[0083] The data preprocessing module 11 is used to preprocess the navigation data to obtain the navigation data image;
[0084] A feature extraction module 12 is used to extract features from the navigation data image based on group convolution, reverse bottleneck structure, SE attention mechanism and feature pyramid structure to obtain a feature map;
[0085] A suggestion box acquisition module 13 is used to input the feature map into the RPN region proposal network to locate the target table and obtain a suggestion box;
[0086] A suggested feature map acquisition module 14 is used to obtain a suggested feature map using a cubic convolution interpolation algorithm according to the feature map and the suggested box;
[0087] The table positioning module 15 is used to perform bounding box regression processing on the proposed feature map to obtain the table area of the navigation data picture.
[0088] Optionally, the feature extraction module is specifically used for:
[0089] Extracting a first feature and a first feature map from the navigation data image based on grouped convolution and reverse bottleneck structure;
[0090] Combining the SE attention mechanism and the feature pyramid structure, the first feature is upsampled to obtain the second feature, and the first feature map is convolved to increase the number of channels to the same as the number of channels of the convolution kernel to obtain the second feature map, and then the second feature and the second feature map are directly added;
[0091] After repeating the above steps several times, several feature maps of different scales are obtained.
[0092] Optionally, the formula of the SE attention mechanism is:
[0093] out=σ(W2δ(W1in))
[0094] Among them, out represents the obtained feature weights of each channel, in represents the feature map of the input attention; σ() represents the Sigmoid activation function, δ() represents the ReLu activation function, W1 represents the parameters of the first fully connected layer, and W2 represents the parameters of the second fully connected layer.
[0095] Optionally, the feature pyramid structure includes a bottom-up path, a top-down path and a lateral connection.
[0096] Optionally, the suggestion box acquisition module is specifically used to:
[0097] Input the feature map into the RPN region proposal network, classify the anchors of the feature map by a normalized exponential function based on the RPN region proposal network, obtain positive anchors with targets and negative anchors without targets, and calculate the bounding box regression offset of the corresponding anchors;
[0098] All suggestions are obtained according to the bounding box regression offsets of the positive anchor and the corresponding anchor, and suggestions that do not meet the preset boundary standards are removed to obtain a suggestion box.
[0099] Optionally, the formula of the cubic convolution interpolation algorithm is:
[0100]
[0101] Among them, g(x) represents the interpolation result of the interpolation function at point x, u represents the unit step function, h represents the sampling interval, x represents the position of the interpolation point, k represents the convolution kernel used for interpolation convolution, and x k represents the position of the corresponding interpolation point in the convolution kernel k, c represents the convolution coefficient, c k Represents the coefficient corresponding to the convolution kernel k.
[0102] Optionally, the navigation data table positioning system further includes:
[0103] The suggested feature graph category determination module is used to perform category determination on the suggested feature graph to obtain the category of the suggested feature graph as the category of the table area.
[0104] It is worth noting that the working process of each module in the navigation data table positioning system 10 described in the embodiment of the present invention can refer to the working process of the navigation data table positioning method described in the above embodiment, and will not be repeated here.
[0105] The navigation data table positioning system provided by the embodiment of the present invention first pre-processes the navigation data to obtain a navigation data picture as input; then uses a ConvNeXt-XL model based on grouped convolution, reverse bottleneck structure, SE attention mechanism and feature pyramid structure to extract features from the navigation data picture to obtain a feature map; inputs the feature map into the RPN region proposal network to locate the target table to obtain a proposal box; then, in the ROI Conv layer, a cubic convolution interpolation algorithm is used to obtain a proposal feature map based on the feature map and the proposal box; finally, a bounding box regression process is performed on the proposal feature map to obtain a table area of the navigation data picture. The embodiment of the present invention extracts multi-size features of the navigation data image for training based on grouped convolution, reverse bottleneck structure, SE attention mechanism and feature pyramid structure. At the same time, the violent rounding method of the ROI pooling layer in the traditional Faster R-CNN model is abandoned, and the cubic convolution interpolation method is introduced for feature mapping, which can produce very smooth edges and avoid the loss of feature map pixels caused by forced rounding in the feature mapping stage, thereby achieving more full utilization of image features, obtaining a more accurate feature extraction frame, and improving the accuracy of navigation data table positioning.
[0106] In addition, an embodiment of the present invention further provides a computer-readable storage medium, which includes a stored computer program; wherein, when the computer program is running, it controls the device where the computer-readable storage medium is located to execute the above-mentioned navigation data table positioning method.
[0107] See also Figure 8 , Figure 8: is a structural block diagram of a navigation data table positioning device 20 provided in an embodiment of the present invention, and the navigation data table positioning device 20 includes: a processor 21, a memory 22, and a computer program stored in the memory 22 and executable on the processor 21. When the processor 21 executes the computer program, the steps in the above-mentioned navigation data table positioning method embodiment are implemented. Alternatively, when the processor 21 executes the computer program, the functions of each module / unit in the above-mentioned system embodiments are implemented.
[0108] Exemplarily, the computer program may be divided into one or more modules / units, which are stored in the memory 22 and executed by the processor 21 to implement the present invention. The one or more modules / units may be a series of computer program instruction segments capable of implementing specific functions, which are used to describe the execution process of the computer program in the navigation data table positioning device 20.
[0109] The navigation data table location device 20 may include, but is not limited to, a processor 21 and a memory 22. Those skilled in the art will appreciate that the schematic diagram is merely an example of the navigation data table location device 20 and does not constitute a limitation on the navigation data table location device 20. The navigation data table location device 20 may include more or fewer components than shown in the diagram, or may combine certain components, or different components. For example, the navigation data table location device 20 may also include input and output devices, network access devices, buses, etc.
[0110] The processor 21 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc. The processor 21 is the control center of the navigation data table positioning device 20, and uses various interfaces and lines to connect various parts of the entire navigation data table positioning device 20.
[0111] The memory 22 can be used to store the computer program and / or module. The processor 21 realizes various functions of the navigation data table positioning device 20 by running or executing the computer program and / or module stored in the memory 22 and calling the data stored in the memory 22. The memory 22 can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area can store data created according to the use of the mobile phone (such as audio data, a phone book, etc.), etc. In addition, the memory 22 can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.
[0112] Wherein, if the module / unit integrated in the navigation data table positioning device 20 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 such an understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by the processor 21, the steps of the above-mentioned method embodiments can be implemented. Wherein, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or system that can carry the computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium.
[0113] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.
Claims
1. A method for locating a navigation data table, characterized in that: include: Preprocess the navigation data to obtain a navigation data image; Based on group convolution, reverse bottleneck structure, SE attention mechanism and feature pyramid structure, feature extraction is performed on the navigation data image to obtain a feature map; The feature map is input into the RPN region proposal network to locate the target table and obtain a proposal box; According to the feature map and the suggestion box, a cubic convolution interpolation algorithm is used to obtain a suggestion feature map; Performing bounding box regression processing on the proposed feature map to obtain a table area of the navigation data image; The method of extracting features from the navigation data image based on group convolution, reverse bottleneck structure, SE attention mechanism and feature pyramid structure to obtain a feature map includes: Extracting a first feature and a first feature map from the navigation data image based on grouped convolution and reverse bottleneck structure; Combining the SE attention mechanism and the feature pyramid structure, the first feature is upsampled to obtain the second feature, and the first feature map is convolved to increase the number of channels to the same as the number of channels of the convolution kernel to obtain the second feature map, and then the second feature and the second feature map are directly added; After repeating the above steps several times, several feature maps of different scales are obtained.
2. The method for locating the navigation data table according to claim 1, characterized in that: The formula of the SE attention mechanism is: out=σ(W2δ(W1in)) Among them, out represents the obtained feature weights of each channel, in represents the feature map of the input attention; σ() represents the Sigmoid activation function, δ() represents the ReLu activation function, W1 represents the parameters of the first fully connected layer, and W2 represents the parameters of the second fully connected layer.
3. The method for locating the navigation data table according to claim 1, characterized in that: The feature pyramid structure includes a bottom-up path, a top-down path and lateral connections.
4. The method for locating a navigation data table according to claim 1, wherein: The step of inputting the feature map into the RPN region proposal network to locate the target table and obtain the proposal box includes: Input the feature map into the RPN region proposal network, classify the anchors of the feature map by a normalized exponential function based on the RPN region proposal network, obtain positive anchors with targets and negative anchors without targets, and calculate the bounding box regression offset of the corresponding anchors; All suggestions are obtained according to the bounding box regression offsets of the positive anchor and the corresponding anchor, and suggestions that do not meet the preset boundary standards are removed to obtain a suggestion box.
5. The method for locating the navigation data table according to claim 1, characterized in that: The formula of the cubic convolution interpolation algorithm is: Among them, g(x) represents the interpolation result of the interpolation function at point x, u represents the unit step function, h represents the sampling interval, x represents the position of the interpolation point, k represents the convolution kernel used for interpolation convolution, and x k represents the position of the corresponding interpolation point in the convolution kernel k, c represents the convolution coefficient, c k Represents the coefficient corresponding to the convolution kernel k.
6. The method for locating the navigation data table according to claim 1, characterized in that: After obtaining the suggested feature map by using a cubic convolution interpolation algorithm according to the feature map and the suggestion box, the navigation data table positioning method further includes: A category determination is performed on the suggested feature map to obtain a category of the suggested feature map to serve as a category of the table area.
7. A navigation data table positioning system, characterized in that: include: A data preprocessing module is used to preprocess the navigation data to obtain a navigation data image; A feature extraction module, used for extracting features from the navigation data image based on grouped convolution, reverse bottleneck structure, SE attention mechanism and feature pyramid structure to obtain a feature map; A suggestion box acquisition module is used to input the feature map into the RPN region proposal network to locate the target table and obtain a suggestion box; A suggested feature map acquisition module, used to obtain a suggested feature map by using a cubic convolution interpolation algorithm according to the feature map and the suggested box; A table positioning module, used for performing bounding box regression processing on the proposed feature map to obtain a table area of the navigation data image; Wherein, the feature extraction module is also used for: Extracting a first feature and a first feature map from the navigation data image based on grouped convolution and reverse bottleneck structure; Combining the SE attention mechanism and the feature pyramid structure, the first feature is upsampled to obtain the second feature, and the first feature map is convolved to increase the number of channels to the same as the number of channels of the convolution kernel to obtain the second feature map, and then the second feature and the second feature map are directly added; After repeating the above steps several times, several feature maps of different scales are obtained.
8. A navigation data table positioning device, characterized in that: The method comprises a processor, a memory and a computer program stored in the memory and configured to be executed by the processor, wherein the processor implements the navigation data table positioning method according to any one of claims 1 to 6 when executing the computer program.
9. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored computer program; wherein, when the computer program is run, it controls the device where the computer-readable storage medium is located to execute the navigation data table positioning method according to any one of claims 1 to 6.
