Digital identification method and device based on residual network, and medium

Abstract

The invention discloses a digital recognition method and device based on a residual network and a medium, and relates to the technical field of intelligent operation and maintenance of rail transit. The method comprises the following steps: based on a digital data set, adopting a residual network method, and obtaining a digital recognition model through network training; and acquiring a real-time image, and inputting each region-of-interest image into the number recognition model based on a preset number recognition region-of-interest to obtain a number value result. The method can adapt to image quality, instrument display diversity and environmental noise under different illumination conditions, is efficient and high in robustness, can realize automatic identification of nixie tube display content, and further improves the accuracy and efficiency of data reading.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent operation and maintenance of rail transit, and more specifically, to a digital recognition method, device and medium based on a residual network. Background Art

[0002] With the rapid development of modern railway transportation, the railway water supply pump room, as one of the important infrastructures, has gradually improved its intelligent and automated operation and maintenance level. In the railway water supply pump room, digital tube instruments are widely used to monitor and display key parameters such as the working status, flow rate, pressure, etc. of the water pump in real time. The accurate reading of these instruments is crucial to ensure the normal operation of the pump room, prevent equipment failures, and improve the overall efficiency of the system.

[0003] The traditional method of instrument reading mainly relies on manual observation and recording, which has the following problems: manual reading errors, high labor intensity, and difficulty in real-time monitoring. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a digital recognition method, device and medium based on a residual network. By using a high-performance camera and a deep learning algorithm, automatic recognition of the content displayed on a digital tube can be achieved, thereby improving the accuracy and efficiency of data reading.

[0005] In a first aspect, the present invention provides a digital recognition method based on a residual network, the method comprising:

[0006] Based on the digital data set, the residual network method is adopted to obtain the digital recognition model through network training;

[0007] Acquire real-time images, and based on the preset digitally recognized regions of interest, input each region of interest image into the digital recognition model to obtain digital numerical results.

[0008] Furthermore, based on the digital data set, a residual network method is adopted to obtain a digital recognition model through network training, including:

[0009] Construct a residual neural network structure;

[0010] The digital data set is imported into the residual neural network structure for training to obtain a digital recognition model.

[0011] Furthermore, the residual neural network structure is based on the ResNet34 structure, including 1 convolutional layer, 4 residual blocks, 2 pooling layers, and 1 fully connected layer; wherein the convolutional layer is used to extract image features; the residual block contains convolution, regularization, activation, and residual connection structures; the pooling layer is used to realize average pooling or maximum pooling; and the fully connected layer is used to output the probability of each result.

[0012] Furthermore, the residual neural network structure includes seven modules, namely a first module, a second module, a third module, a fourth module, a fifth module, a sixth module and a seventh module; wherein:

[0013] The first module is a convolutional layer with a kernel size of 7×7 and a stride of 2, followed by a maximum pooling layer with a size of 2×2 and a stride of 2, with an output size of 56×56×64;

[0014] The second module is a residual block, which includes three convolution blocks. Each convolution block is two convolution layers of size 3×3 and stride 1. The input and output are added for residual connection, and the output size is 56×56×64.

[0015] The third module is a residual block, which includes 4 convolution blocks, each of which is 2 convolution layers of size 3×3. The input is added to the output for residual connection after adjusting the number of channels through 1×1 convolution, and the output size is 28×28×128.

[0016] The fourth module is a residual block, which includes 6 convolution blocks, each of which is 2 convolution layers of size 3×3. The input is added to the output for residual connection after adjusting the number of channels through 1×1 convolution, and the output size is 14×14×256.

[0017] The fifth module is a residual block, which includes 3 convolution blocks, each of which is 2 convolution layers of size 3×3. The input is added to the output for residual connection after adjusting the number of channels through 1×1 convolution, and the output size is 7×7×256.

[0018] The sixth module is an average pooling layer, and the output size is 1×1×128;

[0019] The seventh module is a fully connected layer, and the number of output channels is 11.

[0020] Furthermore, the digital data set is imported into the residual neural network structure for training to obtain a digital recognition model, including:

[0021] Inputting the digital data set into the residual neural network structure to obtain a feature map until the fully connected layer obtains a prediction result;

[0022] Based on the prediction results, the residual neural network structure is trained using a loss function, so that the trained residual neural network structure is used as a digital recognition model.

[0023] Furthermore, the prediction result is a 1×11 vector, which respectively represents the probability values ​​of the detection result being a number from 0 to 9 and a decimal point.

[0024] Furthermore, the loss function is expressed as:

[0025]

[0026] Where L is the loss value, p i is the output probability of each label; C is the set of all labels; a i and superscript r are hyperparameters determined by the sample distribution; N is the regularization term, which indicates the complexity of the model parameters; λ is the proportion of the regularization term to the total loss.

[0027] Furthermore, a real-time image is acquired, and based on the preset digitally recognized regions of interest, each region of interest image is input into a digital recognition model to obtain a digital numerical result, including:

[0028] Set the regions of interest for each single digit and decimal point in the recognized digits, and input them into the digit recognition model in sequence to obtain single digit labels;

[0029] Each digital label is concatenated in sequence to obtain a string result of the complete value.

[0030] In a second aspect, the present invention provides a digital recognition device based on a residual network, the device comprising:

[0031] The model building unit is configured to obtain a digital recognition model through network training using a residual network method based on a digital data set;

[0032] The numerical recognition unit is configured to acquire real-time images, input each region of interest image into the digital recognition model based on the preset region of interest for digital recognition, and obtain a digital numerical result.

[0033] In a third aspect, the present invention provides a readable storage medium, wherein the readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the method as described above.

[0034] The present invention has at least the following beneficial effects:

[0035] The present invention can adapt to image quality under different lighting conditions, the diversity of instrument displays, and environmental noise. The method is efficient and highly robust, and can realize automatic recognition of digital tube display content, thereby improving the accuracy and efficiency of data reading. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A flowchart of a residual network-based digit recognition method according to an embodiment of the present invention is shown.

[0037] Figure 2A flowchart for constructing a digital recognition model according to an embodiment of the present invention is shown.

[0038] Figure 3 A training flow chart of a residual neural network structure according to an embodiment of the present invention is shown.

[0039] Figure 4 A flowchart of digital recognition according to an embodiment of the present invention is shown.

[0040] Figure 5 The output result of the digital recognition model according to the embodiment of the present invention is shown.

[0041] Figure 6 A structural diagram of a digital recognition device based on a residual network according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0042] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. The embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings and specific embodiments, but are not intended to limit the present invention. For the various steps described herein, if there is no necessity for a causal relationship between each other, the order in which they are described as examples herein should not be regarded as a limitation, and those skilled in the art should know that they can be adjusted in order, as long as the logic between them is not destroyed, resulting in the inability to implement the entire process.

[0043] The embodiment of the present invention provides a digital recognition method based on residual network, such as Figure 1 FIG. 1 is a flowchart of a digital recognition method based on a residual network. The digital recognition method based on a residual network includes steps S10 to S20, which are described in detail as follows.

[0044] S10: Based on the digital data set, the residual network method is used to obtain the digital recognition model through network training.

[0045] In this embodiment, the digital data set selects a track data set, which consists of pictures and labels of single digits, the labels are the numerical values ​​of the digits, the labels of the digits 0 to 9 are 0 to 9 respectively, and the label of the decimal point is 10; in this embodiment, there are 260,000 pictures and labels, the training set ratio is 80%, and the test set ratio is 20%.

[0046] In some embodiments, Figure 2 As shown in FIG. 1 , it is a flowchart for constructing a digital recognition model, and step S10 is specifically implemented by the following steps:

[0047] S11: Construct a residual neural network structure.

[0048] In this embodiment, the residual neural network structure is based on the ResNet34 structure, including 1 convolutional layer, 4 residual blocks, 2 pooling layers, and 1 fully connected layer; wherein the convolutional layer is used to extract image features; the residual block contains convolution, regularization, activation, and residual connection structures; the pooling layer is used to realize average pooling or maximum pooling; and the fully connected layer is used to output the probability of each result.

[0049] Exemplarily, the residual neural network structure includes seven modules, namely, a first module, a second module, a third module, a fourth module, a fifth module, a sixth module and a seventh module; wherein:

[0050] The first module is a convolutional layer with a kernel size of 7×7 and a stride of 2, followed by a maximum pooling layer with a size of 2×2 and a stride of 2, with an output size of 56×56×64;

[0051] The second module is a residual block, which contains 3 convolution blocks. Each convolution block is 2 convolution layers of size 3×3 and stride 1. The input and output are added for residual connection, and the output size is 56×56×64.

[0052] The third module is the residual block, which contains 4 convolution blocks. Each convolution block is 2 convolution layers of size 3×3. The input is added to the output for residual connection after adjusting the number of channels through 1×1 convolution. The output size is 28×28×128.

[0053] The fourth module is the residual block, which contains 6 convolution blocks. Each convolution block is 2 convolution layers of size 3×3. The input is added to the output for residual connection after adjusting the number of channels through 1×1 convolution. The output size is 14×14×256.

[0054] The fifth module is a residual block, which contains three convolution blocks. Each convolution block is two 3×3 convolution layers. The input is added to the output for residual connection after adjusting the number of channels through 1×1 convolution. The output size is 7×7×256.

[0055] The sixth module is the average pooling layer with an output size of 1×1×128;

[0056] The seventh module is a fully connected layer with 11 output channels.

[0057] S12: Import the digital data set into the residual neural network structure for training to obtain a digital recognition model.

[0058] In some embodiments, Figure 3 As shown, it is a training flow chart of the residual neural network structure, and step S12 is specifically implemented by the following steps:

[0059] S121: Input the digital data set into the residual neural network structure to obtain a feature map until the fully connected layer obtains a prediction result;

[0060] In this embodiment, the prediction result is a 1×11 vector, which respectively represents the probability value of the detection result being a number from 0 to 9 and a decimal point.

[0061] S122: Based on the prediction results, the residual neural network structure is trained using the loss function, and the trained residual neural network structure is used as a digital recognition model.

[0062] In this embodiment, the loss function is expressed as:

[0063]

[0064] Where L is the loss value, p i is the output probability of each label; C is the set of all labels; a i and superscript r are hyperparameters determined by the sample distribution; N is the regularization term, which indicates the complexity of the model parameters; λ is the proportion of the regularization term to the total loss.

[0065] For example, the ratio of the regularization term to the total loss is 0.2; the hyperparameter a i The values ​​of and superscript r are 0.25 and 2, respectively. When using the loss function to train the residual neural network structure, the Adam optimizer is used for network training, the predetermined number of iterations is 400 steps, and the learning rate is 0.00015.

[0066] S20: Acquire a real-time image, and based on the preset digitally recognized regions of interest, input each region of interest image into a digital recognition model to obtain a digital numerical result.

[0067] In some embodiments, Figure 4 As shown in the figure, it is a digital recognition flow chart, and step S20 is specifically implemented by the following steps:

[0068] S21: Setting the regions of interest for each single digit and decimal point in the recognized digits, and inputting them into the digit recognition model in sequence to obtain single digit labels.

[0069] S22: The digital labels are concatenated in sequence to obtain a string result of a complete numerical value.

[0070] For example, the output of the digital recognition model is as follows: Figure 5 As shown in the figure, the image size is 420*420, 5 regions of interest are set, and the model output labels are "0", ".", "2", "4", and "8" respectively; the final numerical result is 0.248.

[0071] The embodiment of the present invention also provides a digital recognition device based on a residual network, such as Figure 6 As shown, the device comprises:

[0072] The model building unit 601 is configured to obtain a digital recognition model through network training using a residual network method based on a digital data set;

[0073] The numerical value recognition unit 602 is configured to acquire real-time images, input each region of interest image into a digital recognition model based on a preset region of interest for digital recognition, and obtain a digital numerical value result.

[0074] In some embodiments, the model building unit is further configured to:

[0075] Construct a residual neural network structure;

[0076] The digital data set is imported into the residual neural network structure for training to obtain a digital recognition model.

[0077] In some embodiments, the residual neural network structure is based on the ResNet34 structure, including 1 convolutional layer, 4 residual blocks, 2 pooling layers, and 1 fully connected layer; wherein the convolutional layer is used to extract image features; the residual block contains convolution, regularization, activation, and residual connection structures; the pooling layer is used to implement average pooling or maximum pooling; the fully connected layer is used to output the probability of each result.

[0078] In some embodiments, the residual neural network structure includes seven modules, namely, a first module, a second module, a third module, a fourth module, a fifth module, a sixth module and a seventh module; wherein:

[0079] The first module is a convolutional layer with a kernel size of 7×7 and a stride of 2, followed by a maximum pooling layer with a size of 2×2 and a stride of 2, with an output size of 56×56×64;

[0080] The second module is a residual block, which includes three convolution blocks. Each convolution block is two convolution layers of size 3×3 and stride 1. The input and output are added for residual connection, and the output size is 56×56×64.

[0081] The third module is a residual block, which includes 4 convolution blocks, each of which is 2 convolution layers of size 3×3. The input is added to the output for residual connection after adjusting the number of channels through 1×1 convolution, and the output size is 28×28×128.

[0082] The fourth module is a residual block, which includes 6 convolution blocks, each of which is 2 convolution layers of size 3×3. The input is added to the output for residual connection after adjusting the number of channels through 1×1 convolution, and the output size is 14×14×256.

[0083] The fifth module is a residual block, which includes 3 convolution blocks, each of which is 2 convolution layers of size 3×3. The input is added to the output for residual connection after adjusting the number of channels through 1×1 convolution, and the output size is 7×7×256.

[0084] The sixth module is an average pooling layer, and the output size is 1×1×128;

[0085] The seventh module is a fully connected layer, and the number of output channels is 11.

[0086] In some embodiments, the model building unit is further configured to:

[0087] Inputting the digital data set into the residual neural network structure to obtain a feature map until the fully connected layer obtains a prediction result;

[0088] Based on the prediction results, the residual neural network structure is trained using a loss function, so that the trained residual neural network structure is used as a digital recognition model.

[0089] In some embodiments, the prediction result is a 1×11 vector, which respectively represents the probability values ​​of the detection result being a number from 0 to 9 and a decimal point.

[0090] In some embodiments, the loss function is expressed as:

[0091]

[0092] Where L is the loss value, p i is the output probability of each label; C is the set of all labels; a i and superscript r are hyperparameters determined by the sample distribution; N is the regularization term, which indicates the complexity of the model parameters; λ is the proportion of the regularization term to the total loss.

[0093] In some embodiments, the value identification unit is further configured to:

[0094] Set the regions of interest for each single digit and decimal point in the recognized digits, and input them into the digit recognition model in sequence to obtain single digit labels;

[0095] Each digital label is concatenated in sequence to obtain a string result of the complete value.

[0096] It should be noted that the structures of the various residual network-based digital recognition devices described in this embodiment belong to the same technical concept as the previously described residual network-based digital recognition method, and achieve the same beneficial effects through the same principles, which will not be repeated here.

[0097] An embodiment of the present invention further provides a readable storage medium, which stores one or more programs. The one or more programs can be executed by one or more processors to implement the method described in any of the above embodiments.

[0098] In addition, although exemplary embodiments have been described herein, the scope includes any and all embodiments based on the present invention with equivalent elements, modifications, omissions, combinations (e.g., various embodiments intersecting schemes), adaptations or changes. The elements in the claims will be interpreted broadly based on the language adopted in the claims, and are not limited to the examples described in this specification or during the implementation of this application, and the examples will be interpreted as non-exclusive. Therefore, this specification and examples are intended to be considered as examples only, and the true scope and spirit are indicated by the following claims and the full scope of their equivalents.

[0099] The above description is intended to be illustrative rather than restrictive. For example, the above examples (or one or more of them) can be used in combination with each other. For example, those of ordinary skill in the art can use other embodiments when reading the above description. In addition, in the above-mentioned specific embodiments, various features can be grouped together to simplify the present invention. This should not be interpreted as a feature of an invention that is not claimed for protection being necessary for any claim. On the contrary, the subject matter of the present invention may be less than all the features of the embodiments of a specific invention. Thus, the following claims are incorporated into the specific embodiments as examples or embodiments, wherein each claim is independently used as a separate embodiment, and it is considered that these embodiments can be combined with each other in various combinations or arrangements. The scope of the present invention should be determined with reference to the full scope of the equivalent forms of the attached claims and these claims.

Claims

1. A digital recognition method based on residual network, characterized in that: The method comprises: Based on the digital data set, the residual network method is adopted to obtain the digital recognition model through network training; Acquire real-time images, and based on the preset digitally recognized regions of interest, input each region of interest image into the digital recognition model to obtain digital numerical results.

2. The residual network-based digital recognition method according to claim 1, characterized in that: Based on the digital data set, the residual network method is used to obtain the digital recognition model through network training, including: Construct a residual neural network structure; The digital data set is imported into the residual neural network structure for training to obtain a digital recognition model.

3. The residual network-based digital recognition method according to claim 2, characterized in that: The residual neural network structure is based on the ResNet34 structure, including 1 convolution layer, 4 residual blocks, 2 pooling layers, and 1 fully connected layer; wherein the convolution layer is used to extract image features; the residual block contains convolution, regularization, activation, and residual connection structures; the pooling layer is used to realize average pooling or maximum pooling; and the fully connected layer is used to output the probability of each result.

4. The residual network-based digital recognition method according to claim 2, characterized in that: The residual neural network structure includes seven modules, namely, a first module, a second module, a third module, a fourth module, a fifth module, a sixth module and a seventh module; wherein: The first module is a convolutional layer with a kernel size of 7×7 and a stride of 2, followed by a maximum pooling layer with a size of 2×2 and a stride of 2, with an output size of 56×56×64; The second module is a residual block, which includes three convolution blocks. Each convolution block is two convolution layers of size 3×3 and stride 1. The input and output are added for residual connection, and the output size is 56×56×64. The third module is a residual block, which includes 4 convolution blocks, each of which is 2 convolution layers of size 3×3. The input is added to the output for residual connection after adjusting the number of channels through 1×1 convolution, and the output size is 28×28×128. The fourth module is a residual block, which includes 6 convolution blocks, each of which is 2 convolution layers of size 3×3. The input is added to the output for residual connection after adjusting the number of channels through 1×1 convolution, and the output size is 14×14×256. The fifth module is a residual block, which includes 3 convolution blocks, each of which is 2 convolution layers of size 3×3. The input is added to the output for residual connection after adjusting the number of channels through 1×1 convolution, and the output size is 7×7×256. The sixth module is an average pooling layer, and the output size is 1×1×128; The seventh module is a fully connected layer, and the number of output channels is 11.

5. The residual network-based digital recognition method according to claim 4, characterized in that: The digital data set is imported into the residual neural network structure for training to obtain a digital recognition model, including: Inputting the digital data set into the residual neural network structure to obtain a feature map until the fully connected layer obtains a prediction result; Based on the prediction results, the residual neural network structure is trained using a loss function, so that the trained residual neural network structure is used as a digital recognition model.

6. The residual network-based digital recognition method according to claim 5, characterized in that: The prediction result is a 1×11 vector, which respectively represents the probability values ​​of the detection result being a number from 0 to 9 and a decimal point.

7. The residual network-based digital recognition method according to claim 6, characterized in that: The loss function is expressed as: Where L is the loss value, p i is the output probability of each label; C is the set of all labels; a i and superscript r are hyperparameters determined by the sample distribution; N is the regularization term, which indicates the complexity of the model parameters; λ is the proportion of the regularization term to the total loss.

8. The residual network-based digital recognition method according to claim 1, characterized in that: Acquire real-time images, based on the preset digitally recognized regions of interest, input each region of interest image into the digital recognition model to obtain digital numerical results, including: Set the regions of interest for each single digit and decimal point in the recognized digits, and input them into the digit recognition model in sequence to obtain single digit labels; Each digital label is concatenated in sequence to obtain a string result of the complete value.

9. A digital recognition device based on a residual network, characterized in that: The device comprises: The model building unit is configured to obtain a digital recognition model through network training using a residual network method based on a digital data set; The numerical recognition unit is configured to acquire real-time images, input each region of interest image into the digital recognition model based on the preset region of interest for digital recognition, and obtain a digital numerical result. 10 . A non-transitory computer-readable storage medium storing instructions, which, when executed by a processor, perform the method according to claim 1 .

Images