Engine cylinder sleeve character recognition method, neural network model and device

Abstract

The invention relates to the field of automobile manufacturing, in particular to a method, a neural network model and a device for character recognition of an engine cylinder sleeve. Comprising a convolution pooling layer, a shallow layer residual module, a middle layer residual module, a deep layer residual module, a dimension reduction convolution layer, a target detection model, a pooling module, a character code spraying classifier and a character string matching judgment device which are connected in sequence, and convolution kernels in the shallow layer residual module and the shallow layer residual module are large convolution kernels. By expanding a receptive field of a convolution kernel, feature information in a larger range is captured in a shallow neural network, and connection is established between the output of a shallow residual module and the input of a deep residual module, so that the model can better integrate detail features and large-range features of character spraying codes during feature extraction, and the accuracy of feature extraction is improved. And the robustness of the model and the accuracy of character recognition are improved.

Description

Technical Field

[0001] The invention relates to the field of automobile manufacturing, and in particular to a method, a neural network model and a device for character recognition of an engine cylinder sleeve. Background Art

[0002] The engine cylinder liner is the core component of the engine. It provides guidance and sealing for the engine's piston movement, while bearing the high temperature and pressure generated by combustion. Each cylinder of an engine is usually equipped with a matching cylinder liner. The surface of the cylinder liner is an uneven cylinder with a character spray code as its unique identification. When assembling the engine cylinder liner and the engine cylinder, it is necessary to identify the character spray code on the cylinder liner and assemble it to the matching engine cylinder to ensure that the engine can work correctly.

[0003] Due to the uneven surface of the engine cylinder liner, and the soft process used when inkjetting characters on the engine cylinder liner, the characters on different engine cylinder liners may have different sizes, different ink depths, different character stroke thicknesses, different intervals between characters, skewed characters, etc. When the existing neural network model is used to recognize the character codes on the engine cylinder liner, it is prone to errors in character recognition of the engine cylinder liner due to its weak robustness, and ultimately cannot be truly used in the assembly process of the engine cylinder liner. Summary of the invention

[0004] Based on this, the purpose of the present invention is to overcome the defects or shortcomings of the prior art and provide a method and a neural network model for engine cylinder liner character recognition.

[0005] A method for character recognition of an engine cylinder liner comprises the following steps:

[0006] S1: Perform convolution pooling processing on the engine cylinder liner image with character coding and reduce the dimension to generate the first feature map.

[0007] S2: Use a large convolution kernel to perform two feature extractions and one dimension reduction on the first feature map to generate a third feature map with a large receptive field to extract detailed features of the engine cylinder liner image with character inkjet coding.

[0008] The above step S2 expands the receptive field of the convolution kernel, thereby capturing feature information in a larger range in the shallow neural network; specifically, when the character coding in the image is too large or the interval between its characters is too large, if the convolution kernel used by the model for feature extraction is too small, the model only focuses on the detailed features in the image and discards the features of a large area, resulting in the model being unable to correctly extract the features of the character coding. By expanding the size of the convolution kernel, the model can focus on the character coding features of a larger area, thereby improving the model's feature extraction capabilities for these different situations, thereby improving the robustness of the model and the accuracy of character recognition.

[0009] S3: extracting features and reducing the dimension of the third feature map, and generating a fourth feature map to extract features of a larger range of the engine cylinder liner image with character inkjet coding.

[0010] S4: Fusing the third feature map with the fourth feature map, extracting features and reducing the dimension of the fused fifth feature map to generate a sixth feature map, so as to extract features of a larger range of the engine cylinder liner image with character inkjet coding.

[0011] S5: Generate an area of ​​interest for the sixth feature map, so as to obtain an area of ​​interest of the character coding in the engine cylinder liner image with the character coding, so that the sorting information of the character coding can be determined through the area of ​​interest of the character coding.

[0012] S6: performing a pooling operation on the sixth feature map and the region of interest of the character coding, so as to obtain a seventh feature map having features of both the location information of the character coding and the features of the sixth feature map.

[0013] S7: Classify and identify the seventh feature map, classify and identify the letters or numbers corresponding to each character in the character coding, and sort the characters according to their position information to obtain the character string corresponding to the character coding in the engine cylinder liner image with the character coding.

[0014] S8: judging whether the character coding in the engine cylinder liner image with character coding matches the engine cylinder according to the character string.

[0015] Furthermore, the convolution kernel size in step S2 is 5×5, so as to expand the receptive field of the convolution kernel in the shallow neural network, so as to capture feature information in a wider range in the shallow neural network.

[0016] Furthermore, the step S4 also includes steps S41 and S42:

[0017] S41: Performing dimensionality reduction processing on the third feature map to form a third feature map after dimensionality reduction, so that the third feature map has the same dimension as the fourth feature map, and concatenating the third feature map after dimensionality reduction with the fourth feature map to form a fifth feature map.

[0018] S42: performing feature extraction and dimensionality reduction on the fifth feature map, and generating a sixth feature map to extract features of a larger range of the engine cylinder liner image with character inkjet coding.

[0019] Through steps S41 and S42, the third feature map with character coding detail features extracted by the model in S2 can be directly passed to step S4 for feature extraction and generation of the sixth feature map, so that the model can better integrate the detail features and large-scale features of the character coding during feature extraction. When faced with changes in the size, ink depth and character spacing of the character coding on the engine cylinder liner, the model will not lose the large-scale position information features of the character coding and the detail features of the letters or numbers corresponding to the character coding, thereby improving the robustness of the model and the accuracy of recognition.

[0020] Further, the module for performing feature extraction or dimensionality reduction on the feature map in steps S2, S3, and S4 is a residual module, and the residual module is composed of a first residual processing layer and a second residual processing layer, wherein the first residual processing layer and the second residual processing layer are both composed of a first residual convolution layer, a second residual convolution layer, and a third residual convolution layer, the first residual convolution layer, the second residual convolution layer, and the third residual convolution layer are sequentially connected, and the input feature map of the first residual convolution layer is also added to the output feature map of the third residual convolution layer by a jump connection to jointly constitute the output feature map of the residual processing layer, and the residual processing layer is used to perform feature extraction and dimensionality reduction on the feature map of the input residual module;

[0021] The output of the first residual processing layer is used as the input of the second residual processing layer, wherein the specific calculation expression of each residual processing layer is as follows:

[0022] y 1 =ReLU(BN(Conv 1×1,p (x)))

[0023] Among them, Conv 1×1,pIndicates that a convolution process with a step length of p is performed on the input x of the residual processing layer using a convolution kernel with a size of 1×1, so as to extract features from the feature map of the input x, wherein the step length p is 1 by default; in the second feature extraction of step S2, step S3, and step S4, the step length p in the convolution kernel of the first residual convolution layer in the first residual processing layer is set to 2, that is, the input x is subjected to dimensionality reduction processing, so as to expand the receptive field when performing feature extraction in the second feature extraction of step S2, step S3, and step S4, and obtain more global feature information;

[0024] The calculation expression of the second residual convolution layer is:

[0025] y 2 =ReLU(BN(Conv i×i,1 (y 1 )))

[0026] Among them, Conv i×i Represents the feature map y output by the first residual convolution layer using a convolution kernel of size i×i 1 Perform a convolution with a stride of 1 on the feature map y output from the first residual convolution layer 1 Extract features, wherein the size of the convolution kernel is 3×3 by default, but in step S2, the size of the convolution kernel is 5×5 to expand the receptive field of the convolution kernel, thereby capturing feature information in a larger range in a shallow neural network;

[0027] The calculation expression of the third residual convolution layer is:

[0028] y 3 =BN(Conv 1×1,1 (y 2 ))

[0029] Among them, Conv 1×1 Represents the feature map y output by the second residual convolution layer using a convolution kernel of size 1×1 2 Perform a convolution with a stride of 1 on the feature map y output from the second residual convolution layer 2 Extract features from

[0030] The input feature map x of the first residual convolution layer is connected to the output feature map y of the third residual convolution layer by a skip connection. 3 The sum of the output feature map y of the residual processing layer is formed 4 The calculation expression is:

[0031] x skip =Conv 1×1,p (x)

[0032] xskip =BN(x skip )

[0033] y 4 =ReLU(y 3 +x skip )

[0034] Among them, Conv 1×1,p It means that the input x of the residual processing layer is convolved with a convolution kernel of size 1×1 with a step length of p to extract features from the feature map of the input x, wherein the step length p is 1 by default, but the step length p in the second feature extraction of step S2, step S3 and step S4 is 2, that is, the input x is subjected to dimensionality reduction processing so that the input feature map x is consistent with the output y of the third residual convolution layer. 3 The dimensions of the two components match so that they can be added smoothly.

[0035] A neural network model for engine cylinder liner character recognition includes a convolution pooling layer, a shallow residual module, a middle residual module, a deep residual module, a target detection model, a pooling module, a character coding classifier and a string matching judgement connected in sequence;

[0036] The convolution kernel of the shallow residual module is a large convolution kernel, which is used to perform two feature extractions and one dimension reduction on the first feature map to generate a third feature map with a large receptive field to extract detail features of the engine cylinder liner image with character inkjet coding;

[0037] By expanding the receptive field of the shallow residual module and the convolution kernel in the shallow residual module, feature information in a larger range can be captured in the shallow neural network; specifically, when the character coding in the image is too large or the interval between its characters is too large, if the convolution kernel used by the model for feature extraction is too small, the model only focuses on the detailed features in the image and discards the features of a large area, resulting in the model being unable to correctly extract the features of the character coding. By expanding the size of the convolution kernel, the model can focus on the character coding features of a larger area, thereby improving the model's feature extraction capabilities for these different situations, thereby improving the robustness of the model and the accuracy of character recognition.

[0038] The middle layer residual module is used to extract features and reduce the dimension of the third feature map, and generate a fourth feature map to extract features of a larger range of the engine cylinder liner image with character spraying;

[0039] The deep residual module is used to perform feature fusion on the third feature map and the fourth feature map, and perform feature extraction and dimension reduction on the fused fifth feature map to generate a sixth feature map, so as to extract features of a larger range of the engine cylinder liner image with character inkjet coding;

[0040] The target detection model is used to generate an area of ​​interest for the sixth feature map, thereby obtaining an area of ​​interest of the character coding in the engine cylinder liner image with the character coding, so that the sorting information of the character coding can be determined by the area of ​​interest of the character coding;

[0041] The pooling module is used to perform a pooling operation on the sixth feature map and the region of interest of the character coding, so as to obtain a seventh feature map having features of both the location information of the character coding and the features of the sixth feature map;

[0042] The character coding classifier is used to classify and identify the seventh feature map, classify and identify the letters or numbers corresponding to each character in the character coding, and sort the characters to obtain a character string corresponding to the character coding in the engine cylinder liner image with the character coding;

[0043] The character string matching judgement device is used to judge whether the character spray code in the engine cylinder liner image with character spray code matches the engine cylinder according to the character string.

[0044] Furthermore, the convolution kernel size in the shallow residual module is 5×5, so as to expand the receptive field of the convolution kernel in the shallow neural network, so as to capture feature information in a wider range in the shallow neural network.

[0045] Further, the residual module is composed of a first residual processing layer and a second residual processing layer, wherein the first residual processing layer and the second residual processing layer are both composed of a first residual convolution layer, a second residual convolution layer and a third residual convolution layer, the first residual convolution layer, the second residual convolution layer and the third residual convolution layer are sequentially connected, and the input feature map of the first residual convolution layer is also added to the output feature map of the third residual convolution layer by a jump connection to form an output feature map of the residual processing layer, and the residual processing layer is used to perform feature extraction and dimensionality reduction processing on the feature map of the input residual module;

[0046] The output of the first residual processing layer is used as the input of the second residual processing layer, wherein the specific calculation expression of each residual processing layer is as follows:

[0047] y 1 =ReLU(BN(Conv 1×1,p (x)))

[0048] Among them, Conv 1×1,pIndicates that a convolution process with a step size of p is performed on the input x of the residual processing layer using a convolution kernel of size 1×1, so as to extract features from the feature map of the input x, wherein the step size p is 1 by default; in the second feature extraction of the shallow residual module, the middle residual module and the deep residual module, the step size p in the convolution kernel of the first residual convolution layer in the first residual processing layer is set to 2, that is, the input x is subjected to dimensionality reduction processing, so as to expand the receptive field when the second feature extraction of the shallow residual module, the middle residual module and the deep residual module perform feature extraction, and obtain more global feature information;

[0049] The calculation expression of the second residual convolution layer is:

[0050] y 2 =ReLU(BN(Conv i×i,1 (y 1 )))

[0051] Among them, Conv i×i Represents the feature map y output by the first residual convolution layer using a convolution kernel of size i×i 1 Perform a convolution with a stride of 1 on the feature map y output from the first residual convolution layer 1 Extract features from the convolution kernel, wherein the size of the convolution kernel is 3×3 by default, but in the shallow residual module, the size of the convolution kernel is 5×5 to expand the receptive field of the convolution kernel, thereby capturing feature information in a larger range in the shallow neural network;

[0052] The calculation expression of the third residual convolution layer is:

[0053] y 3 =BN(Conv 1×1,1 (y 2 ))

[0054] Among them, Conv 1×1 Represents the feature map y output by the second residual convolution layer using a convolution kernel of size 1×1 2 Perform a convolution with a stride of 1 on the feature map y output from the second residual convolution layer 2 Extract features from

[0055] The input feature map x of the first residual convolution layer is connected to the output feature map y of the third residual convolution layer by a skip connection. 3 The sum of the output feature map y of the residual processing layer is formed 4 The calculation expression is:

[0056] x skip =Conv 1×1,p (x)

[0057] x skip =BN(x skip )

[0058] y 4 =ReLU(y 3 +x skip )

[0059] Among them, Conv 1×1,p Indicates that a convolution kernel of size 1×1 is used to perform convolution processing with a step size of p on the input x of the residual processing layer to extract features from the feature map of the input x, wherein the step size p is 1 by default, but the step size p in the second feature extraction of the shallow residual module, the middle residual module and the deep residual module is 2, that is, the input x is subjected to dimensionality reduction processing so that the input feature map x is consistent with the output y of the third residual convolution layer. 3 The dimensions of the two components match so that they can be added smoothly.

[0060] Furthermore, the feature map output by the shallow residual module is reduced to the same dimension as the feature map output by the middle residual module and is spliced ​​with the feature map output by the middle residual module and then input into the deep residual module together. The feature map with detailed features of the character coding extracted by the model in the shallow layer can be directly transmitted to the deep residual module for feature extraction and feature map generation, so that the model can better integrate the detailed features and large-scale features of the character coding during feature extraction. When faced with changes in the size of the character coding, ink depth, and character spacing on the engine cylinder sleeve, the model will not lose the large-scale position information features of the character coding and the detailed features of the letters or numbers corresponding to the character coding, thereby improving the robustness of the model and the accuracy of recognition.

[0061] A device for character recognition of an engine cylinder sleeve, comprising a camera, a neural network model and an assembly device;

[0062] The camera is used to capture the image on the engine cylinder liner to obtain the engine cylinder liner image with character spray code;

[0063] The assembly equipment is used to install the engine cylinder liner on the engine;

[0064] The neural network model is the above-mentioned neural network model; the engine cylinder liner image with character coding is input into the neural network model for feature extraction and recognition classification and outputs a character string corresponding to the character coding on the engine cylinder liner image with character coding, and judges whether the engine cylinder liner matches the engine through the character string; if they match, a signal is transmitted to the assembly equipment to install the engine cylinder liner on the engine; if they do not match, the engine cylinder liner is not installed on the engine.

[0065] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned engine cylinder liner character recognition method or neural network model is implemented.

[0066] For better understanding and implementation, the present invention is described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 A structural diagram of a neural network model for engine cylinder liner character recognition;

[0068] Figure 2 A schematic diagram of a method for character recognition of an engine cylinder liner;

[0069] Figure 3 It is the structure diagram of the residual module;

[0070] Figure 4 This is the structure diagram of the residual processing layer. DETAILED DESCRIPTION

[0071] In order to solve the problem that when the existing neural network model recognizes the character spray code on the engine cylinder sleeve, the recognition error occurs due to the different sizes of the character spray codes on different engine cylinder sleeves, different ink depths, different character stroke thicknesses, different intervals between characters, and skewed characters. The present invention increases the size of the convolution kernel in the neural network to expand the receptive field of the convolution layer when recognizing the characters on the engine cylinder sleeve, so as to capture a wider range of feature information in the engine cylinder sleeve image, thereby improving the robustness of the neural network model for character recognition; at the same time, a connection is added between the shallow and deep layers of the neural network model, and the features extracted by the model in the shallow stage are directly transferred to the deep stage, so that the model can better integrate features at different levels during feature extraction, so that it is easier to learn useful features, thereby improving the expression ability of the model and improving the accuracy of character recognition.

[0072] Based on the above-mentioned design of the character coding recognition scheme on the engine cylinder liner, the present invention designs a device for character recognition of the engine cylinder liner, including a camera, a neural network model and an assembly device; the camera is used to shoot the image on the engine cylinder liner to obtain the engine cylinder liner image with character coding; the assembly device is used to install the engine cylinder liner on the engine; the engine cylinder liner image with character coding can be input into the neural network model for feature extraction and recognition classification, and then the character string corresponding to the character coding on the engine cylinder liner image with character coding is output, and the character string is used to determine whether the engine cylinder liner matches the engine; if it matches, a signal is transmitted to the assembly device to install the engine cylinder liner on the engine; if it does not match, the engine cylinder liner is not installed on the engine.

[0073] For details, please refer to Figure 1 and Figure 2 ,in Figure 1 A neural network model for character recognition of an engine cylinder liner according to the present invention is provided. Figure 2 A method for character recognition of an engine cylinder liner is implemented based on the model. The neural network model for character recognition of an engine cylinder liner comprises a convolution pooling layer 1, a shallow residual module 2, a shallow residual module 3, a deep residual module 4, an RPN network 5, a ROIPooling layer 6, a character coding classifier 7 and a string matching judge 8 connected in sequence.

[0074] The convolution pooling layer 1 is used to execute step S1: reducing the dimension of the image of the engine cylinder liner to generate a first feature map.

[0075] Specifically, the calculation expression for reducing the dimension of the image of the engine cylinder liner is as follows:

[0076] Y 1 =MaxPool 3×3,2 (ReLU(BN(Conv 7×7,2 (X))))

[0077] In the formula, Conv 7×7,2It indicates that a convolution kernel of size 7×7 is used to perform convolution processing with a step length of 2 on the image of the engine cylinder liner with character coding, so as to extract feature information of details from the image X of the engine cylinder liner with character coding, so as to better identify the letters or numbers corresponding to each character in the character coding, and reduce its dimension to a size of 256×256; the image of the engine cylinder liner with character coding is resized to an image X of 512×512 by stretching (Resize); BN represents a normalization operation, so that the input data maintains a more stable distribution, thereby accelerating the training process and convergence speed; ReLU represents an activation function, which is used to introduce nonlinear changes and set the negative values ​​output by the convolution layer to zero, thereby enhancing the significance of positive features; MaxPool 3×3,2 It represents using a 3×3 maximum pooling layer to perform a pooling process with a step size of 2 on the 256×256 size after the previous convolutional layer, and reducing its dimension to a size of 128×128, and finally obtaining the first feature map Y 1 , the first feature map Y 1 The detailed features of the engine cylinder liner image X with character encoding are provided.

[0078] The shallow residual module 2 is used to execute step S2: using a large convolution kernel to perform two feature extractions on the first feature map to generate a third feature map with a large receptive field.

[0079] Specifically, the specific calculation expression for performing two feature extractions on the first feature map using a large convolution kernel is as follows:

[0080] Y 2 = bottleneck p (Y 1 )

[0081] Y 3 = bottleneck p (Y 2 )

[0082] Where Y 2 Represents the first feature map Y 1 Performing the first bottleneck p (Y 1 ) The second feature map with a size of 128×128 after feature extraction; where Y 2 Represents the first feature map Y 1 The third feature map after two feature extractions, that is, the second feature map Y 2 Further bottleneck p (Y 2 ) After feature extraction and dimensionality reduction, the feature map size is 64×64. p(x) represents the residual module, which performs residual processing on the input feature map x, where p represents the adjustable step size; please also refer to Figure 3 and Figure 4 , the residual module bottleneck p (x) is composed of a first residual processing layer and a second residual processing layer, wherein the first residual processing layer and the second residual processing layer are both composed of a first residual convolution layer, a second residual convolution layer and a third residual convolution layer, the first residual convolution layer, the second residual convolution layer and the third residual convolution layer are sequentially connected, and the input feature map of the first residual convolution layer is also added to the output feature map of the third residual convolution layer by a jump connection to form the output feature map of the residual processing layer, and the residual processing layer is used to perform feature extraction and dimensionality reduction processing on the feature map of the input residual module;

[0083] The output of the first residual processing layer is used as the input of the second residual processing layer, wherein the specific calculation expression of each residual processing layer is as follows:

[0084] y 1 =ReLU(BN(Conv 1×1,p (x)))

[0085] Among them, Conv 1×1,p Indicates that a convolution process with a step size of p is performed on the input x of the residual processing layer using a convolution kernel of size 1×1, so as to extract features from the feature map of the input x, wherein the step size p is 1 by default; in the second feature extraction of the shallow residual module 2, the middle residual module and the deep residual module, the step size p in the convolution kernel of the first residual convolution layer in the first residual processing layer is set to 2, that is, the input x is subjected to dimensionality reduction processing, so as to expand the receptive field when the second feature extraction of the shallow residual module 2, the middle residual module and the deep residual module perform feature extraction, and obtain more global feature information;

[0086] The calculation expression of the second residual convolution layer is:

[0087] y 2 =ReLU(BN(Conv i×i,1 (y 1 )))

[0088] Among them, Conv i×i Represents the feature map y output by the first residual convolution layer using a convolution kernel of size i×i 1 Perform a convolution with a stride of 1 on the feature map y output from the first residual convolution layer 1In the invention, features are extracted from the image, wherein the size of the convolution kernel is 3×3 by default, but in the shallow residual module 2, the size of the convolution kernel is 5×5 to expand the receptive field of the convolution kernel, thereby capturing feature information in a larger range in the shallow neural network; specifically, when the character coding in the image is too large or the interval between its characters is too large, if the convolution kernel used by the model for feature extraction is too small, the model only focuses on the detailed features in the image and discards the features of a large area, resulting in the model being unable to correctly extract the features of the character coding. By expanding the size of the convolution kernel, the model can focus on the features of the character coding in a larger area, thereby improving the model's feature extraction capability for these different situations, thereby improving the robustness of the model and the accuracy of character recognition.

[0089] The calculation expression of the third residual convolution layer is:

[0090] y 3 =BN(Conv 1×1,1 (y 2 ))

[0091] Among them, Conv 1×1 Represents the feature map y output by the second residual convolution layer using a convolution kernel of size 1×1 2 Perform a convolution with a stride of 1 on the feature map y output from the second residual convolution layer 2 Extract features from

[0092] The input feature map x of the first residual convolution layer is connected to the output feature map y of the third residual convolution layer by a skip connection. 3 The sum of the output feature map y of the residual processing layer is formed 4 The calculation expression is:

[0093] x skip =Conv 1×1,p (x)

[0094] x skip =BN(x skip )

[0095] y 4 =ReLU(y 3 +x skip )

[0096] Among them, Conv 1×1,pIt means that the input x of the residual processing layer is convolved with a convolution kernel of size 1×1 with a step length of p to extract features from the feature map of the input x, wherein the step length p is 1 by default, but the step length p in the convolution kernel of the first residual convolution layer in the first residual processing layer of the shallow residual module 2, the middle residual module 3 and the deep residual module 4 is 2, that is, the input x is subjected to dimensionality reduction processing so that the input feature map x is consistent with the output y of the third residual convolution layer. 3 The dimensions of the two components match so that they can be added smoothly.

[0097] Accordingly, the present invention uses a shallow residual module 2 with a convolution kernel of size 5×5 to perform feature extraction with a step size of 1 on the first feature map, and generates a second feature map with a size of 128×128 with a large receptive field to extract the detail features of the engine cylinder liner image with character inkjet coding, and further uses a convolution kernel of size 5×5 to perform feature extraction and dimensionality reduction with a step size of 2 on the second feature map with a size of 128×128, and generates a third feature map with a size of 64×64 with a large receptive field to further extract a larger range of detail features of the engine cylinder liner image with character inkjet coding.

[0098] The middle layer residual module 3 is used to perform step S3: 3 Perform feature extraction and dimensionality reduction in sequence to generate the fourth feature map Y with a size of 32×32 4 ; The convolution kernel size in the second residual convolution layer in the middle residual module 3 is 3×3, and the step size p of the first residual convolution layer is 2; to extract a larger range of features of the engine cylinder liner image with character inkjet coding, the specific calculation expression is as follows:

[0099] Y 4 = bottleneck(Y 3 )

[0100] Where Y 4 It is the fourth feature map, and its size is 32×32.

[0101] The deep residual module 4 is used to perform step S4: transform the third feature map Y 3 With the fourth characteristic graph Y 4 Perform feature fusion to generate the fifth feature map Y 5 , and the fifth feature map Y after fusion 5 Perform feature extraction and dimensionality reduction to generate the sixth feature map Y 6 .

[0102] Specifically, a convolution layer with a convolution kernel size of 1×1 is used to convolution the third feature map Y 3 Perform dimensionality reduction with a step size of 2 to form a reduced dimensional third feature map Y with a size of 32×32 3', so that it is consistent with the fourth characteristic map Y 4 The dimension of the third feature map Y 3 ′ and the fourth characteristic graph Y 4 The fifth feature map Y is formed by splicing with a size of 32×32 5 ; Its calculation expression is as follows:

[0103] Y 3 ′=Conv 1×1,2 (Y 3 )

[0104] Y 5 =concat(Y 3 ′,Y 4 )

[0105] In the above formula, Conv 1×1,2 Indicates that the third feature map Y is processed with a convolution kernel of size 1×1 3 Perform convolution with a step size of 2 to form the third feature map Y after dimensionality reduction 3 ′, and the fourth characteristic graph Y 4 The same dimension; concat means that the third feature map Y 3 ' and the fourth characteristic graph Y 4 The channels are spliced ​​to form the fifth feature map Y 5 .

[0106] Next, the fifth feature map Y is processed by the deep residual module 4. 5 Perform feature extraction and dimensionality reduction to generate the sixth feature map Y with a size of 16×16 6 , the convolution kernel size in the second residual convolution layer in the deep residual module 4 is 3×3, and the step size p of the first residual convolution layer is 2; so as to extract a larger range of features of the engine cylinder liner image with character inkjet coding;

[0107] The calculation expression is as follows:

[0108] Y 6 = bottleneck(Y 5 )

[0109] Where Y 6 It is the fourth feature map, and its size is 16×16.

[0110] Accordingly, through feature fusion, the character coding detail features extracted by the model in the shallow residual module 2 can be directly transmitted to the deep residual module 4, so that the model can better integrate the detail features and large-scale features of the character coding during feature extraction. Therefore, when faced with changes in the character coding size, ink depth and character spacing on the engine cylinder liner, the model will not lose the large-scale position information features of the character coding and the detail features of the letters or numbers corresponding to the character coding, thereby improving the robustness of the model and the accuracy of recognition.

[0111] The RPN network 5 is used to perform step S5: 6 Generate the region of interest to obtain the region of interest for character coding.

[0112] Specifically, the sixth characteristic graph Y 6 The ROI region candidate box is generated by the RPN network to identify the possible location proposals of the character coding in the engine cylinder liner image X with character coding, that is, the region of interest of the character coding, so that the sorting information of the character coding can be determined by the location of the character coding in the engine cylinder liner image X with character coding; the calculation expression is as follows:

[0113] Proposals = RPN(Y 6 )

[0114] Wherein, Proposals is the region of interest of character coding, which is used to indicate the possible location of the character coding in the engine cylinder liner image X with character coding; RPN is used to generate a ROI region candidate box to identify the possible location of the character coding in the engine cylinder liner image X with character coding.

[0115] The ROIPooling layer 6 is used to perform step S6: 6 Pooling is performed with the proposals of the region of interest of the character coding to generate the seventh feature map.

[0116] Specifically, for the sixth feature map Y 6 The character coding is pooled in the region of interest of the character coding, thereby obtaining a feature having the location information of the character coding and a sixth feature map Y 6 The seventh characteristic graph of the characteristics of Y 7 ; Its calculation expression is as follows:

[0117] Y 7 =ROIPooling(Proposals,Y 6 )

[0118] In the formula, ROIPooling is used to perform pooling on the input region of interest and feature map.

[0119] The character coding classifier 7 is used to perform step S7: 7 Perform classification and recognition to generate a string String corresponding to the character coding.

[0120] Specifically, for the seventh feature map Y 7 Normalize and calculate the probability distribution, classify and identify the letters or numbers corresponding to each character in the character coding, and sort the characters according to their position information Position, and finally output the string String corresponding to the character coding in the engine cylinder liner image with character coding; its calculation expression is as follows:

[0121] (String, Position)=Softmax(Y 7 )

[0122] Where, String is the string corresponding to the sorted character coding; Positio is the position of the character coding in the engine cylinder liner image X with character coding; Softmax is a Softmax classifier, which is used to normalize the input feature map and calculate the probability distribution of the feature corresponding label.

[0123] The character string matching judgement device 8 is used to execute step S8: installing the engine cylinder and the engine cylinder liner according to the matching relationship between the character string corresponding to the character coding and the engine.

[0124] Specifically, the string String is finally sent to the string matching judgement device to judge whether it matches the cylinder of the engine: if so, the engine cylinder and the engine cylinder liner are installed; if not, the engine cylinder and the engine cylinder liner are not installed.

[0125] Compared with the existing neural network model for engine cylinder liner character recognition, the present invention expands the receptive field of the convolution kernel in the shallow residual module 2 so that when the size or interval of the character coding becomes larger, the model can more easily extract the characteristics of the character coding in the shallow neural network; at the same time, by splicing the third feature map output by the shallow residual module with the fourth feature map output by the middle residual module to form a fifth feature map and then sending it to the deep residual module for residual processing, the model can further perform feature extraction after the detailed features of the character coding extracted by the shallow neural network are fused with the large-scale features of the character coding extracted by the deep neural network, so as not to lose the detailed features and large-scale features of the character coding, thereby improving the accuracy and robustness of the model in recognizing the character coding on the engine cylinder liner image with character coding.

[0126] Based on the same inventive concept, the present application also provides an electronic device, which may be a terminal device such as a server, a desktop computing device or a mobile computing device (e.g., a laptop computing device, a handheld computing device, a tablet computer, a netbook, etc.). The device includes one or more processors and a memory, wherein the processor is used to execute a program to implement the engine cylinder liner character recognition method of the embodiment of the present invention; and the memory is used to store a computer program executable by the processor.

[0127] Based on the same inventive concept, the present application also provides a computer-readable storage medium, corresponding to the aforementioned embodiment of the method for recognizing characters on an engine cylinder liner, wherein the computer-readable storage medium stores a computer program thereon, and when the program is executed by a processor, the steps of the method for recognizing characters on an engine cylinder liner recorded in any of the aforementioned embodiments are implemented.

[0128] The present application may take the form of a computer program product implemented on one or more storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing program code. Computer-usable storage media include permanent and non-permanent, removable and non-removable media, and information storage can be achieved by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include but are not limited to: phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, read-only compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, tape disk storage or other magnetic storage devices or any other non-transmission medium that can be used to store information that can be accessed by a computing device.

[0129] The above-mentioned embodiments only express several implementation methods of the present invention, and the description is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention patent. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present invention, and the present invention is also intended to include these modifications and modifications.

Claims

1. A method for character recognition of an engine cylinder liner, characterized in that: The following steps are involved: S1: performing convolution pooling processing on the image of the engine cylinder liner with character inkjet coding and reducing the dimension to generate a first feature map; S2: Use a large convolution kernel to perform two feature extractions and one dimensionality reduction on the first feature map to generate a third feature map with a large receptive field; S3: extracting features and reducing the dimension of the third feature map, and generating a fourth feature map; S4: performing feature fusion on the third feature map and the fourth feature map, and performing feature extraction and dimension reduction on the fused fifth feature map to generate a sixth feature map; S5: generating an area of ​​interest for the sixth feature map, thereby obtaining an area of ​​interest for the character coding in the engine cylinder liner image with the character coding; S6: performing a pooling operation on the sixth feature map and the region of interest of the character coding, so as to obtain a seventh feature map having features of the location information of the character coding and features of the sixth feature map; S7: classifying and identifying the seventh feature map, classifying and identifying the letters or numbers corresponding to each character in the character coding, and sorting the characters according to their position information to obtain a character string corresponding to the character coding in the engine cylinder liner image with the character coding; S8: judging whether the character coding in the engine cylinder liner image with character coding matches the engine cylinder according to the character string.

2. The method for character recognition of an engine cylinder liner according to claim 1, characterized in that: The convolution kernel size in steps S2 and S3 is 5×5.

3. The method for character recognition of an engine cylinder liner according to claim 2, characterized in that: The step S5 also includes steps S41 and S42: S41: Perform dimensionality reduction processing on the third feature map to form a third feature map after dimensionality reduction, so that the dimension of the third feature map is the same as that of the fourth feature map, and concatenate the third feature map after dimensionality reduction with the fourth feature map to form a fifth feature map, whose calculation expression is: Y3′=Conv 1×1,2 (Y3) Y5=concat(Y3′,Y4) In the formula, Conv 1×1,2 It means that the third feature map Y3 is convolved with a convolution kernel of size 1×1 with a step length of 2 to form a third feature map Y3′ after dimensionality reduction, which has the same dimension as the fourth feature map Y4; concat means that the channels of the third feature map Y3′ and the fourth feature map Y4 are concatenated to form the fifth feature map Y5; S42: Perform feature extraction and dimension reduction on the fifth feature map, and generate a sixth feature map.

4. The method for character recognition of an engine cylinder liner according to claim 3, characterized in that: The module for performing feature extraction or dimensionality reduction on the feature map in steps S2, S3, and S4 is a residual module, and the residual module is composed of a first residual processing layer and a second residual processing layer, wherein the first residual processing layer and the second residual processing layer are both composed of a first residual convolution layer, a second residual convolution layer, and a third residual convolution layer, the first residual convolution layer, the second residual convolution layer, and the third residual convolution layer are sequentially connected, and the input feature map of the first residual convolution layer is also added to the output feature map of the third residual convolution layer by a jump connection to jointly constitute the output feature map of the residual processing layer, and the residual processing layer is used to perform feature extraction and dimensionality reduction on the feature map of the input residual module; The output of the first residual processing layer is used as the input of the second residual processing layer, wherein the specific calculation expression of each residual processing layer is as follows: y1=ReLU(BN(Conv 1×1,p (x))) Among them, Conv 1×1,p Indicates that a convolution process with a step length of p is performed on the input x of the residual processing layer using a convolution kernel with a size of 1×1, wherein the step length p is 1 by default; in the second feature extraction of step S2, step S3 and step S4, the step length p in the convolution kernel of the first residual convolution layer in the first residual processing layer is set to 2, that is, the input x is subjected to dimensionality reduction processing; The calculation expression of the second residual convolution layer is: y2=ReLU(BN(Conv i×i,1 (y1))) Among them, Conv i×i Indicates that a convolution process with a step size of 1 is performed on the feature map y1 output by the first residual convolution layer using a convolution kernel with a size of i×i, wherein the size of the convolution kernel is 3×3 by default, but in the step S2, the size of the convolution kernel is 5×5; The calculation expression of the third residual convolution layer is: <h2 style=";text-align:left;direction:ltr">y3=BN(Conv<h2 style=";text-align:left;direction:ltr"> 1×1,1 <h2 style=";text-align:left;direction:ltr"> (y2)) Among them, Conv 1×1 Indicates that a convolution kernel of size 1×1 is used to perform convolution processing with a step size of 1 on the feature map y2 output by the second residual convolution layer; The input feature map x of the first residual convolution layer is added to the output feature map y3 of the third residual convolution layer by skip connection to form the output feature map y4 of the residual processing layer. The calculation expression is: x skip =Conv 1×1,p (x) x skip =BN(x skip ) y4=ReLU(y3+x skip ) Among them, Conv 1×1,p It means that a convolution process with a step size of p is performed on the input x of the residual processing layer using a convolution kernel of size 1×1, wherein the step size p is 1 by default, but the step size p in the second feature extraction of step S2, step S3 and step S4 is 2, that is, the input x is subjected to dimensionality reduction processing so that the dimension of the input feature map x matches the dimension of the output y3 of the third residual convolution layer so that they can be added smoothly.

5. A neural network model for engine cylinder liner character recognition, characterized in that: It includes sequentially connected convolutional pooling layers, shallow residual modules, middle residual modules, deep residual modules, target detection models, pooling modules, character coding classifiers and string matching judges; The convolution kernel of the shallow residual module is a large convolution kernel, which is used to perform two feature extractions and one dimensionality reduction on the first feature map to generate a third feature map with a large receptive field; The middle layer residual module is used to extract features and reduce the dimension of the third feature map, and generate a fourth feature map; The deep residual module is used to perform feature fusion on the third feature map and the fourth feature map, and perform feature extraction and dimensionality reduction on the fused fifth feature map to generate a sixth feature map; The target detection model is used to generate an area of ​​interest for the sixth feature map, so as to obtain an area of ​​interest of the character coding in the engine cylinder liner image with the character coding; The pooling module is used to perform a pooling operation on the sixth feature map and the region of interest of the character coding, so as to obtain a seventh feature map having features of both the location information of the character coding and the features of the sixth feature map; The character coding classifier is used to classify and identify the seventh feature map, classify and identify the letters or numbers corresponding to each character in the character coding, and sort the characters to obtain a character string corresponding to the character coding in the engine cylinder liner image with the character coding; The character string matching judgement device is used to judge whether the character spray code in the engine cylinder liner image with character spray code matches the engine cylinder according to the character string.

6. The neural network model for engine cylinder liner character recognition according to claim 5 is characterized in that: The shallow residual module and the convolution kernel size in the shallow residual module are 5×5.

7. The neural network model for engine cylinder liner character recognition according to claim 6 is characterized in that: The residual module is composed of a first residual processing layer and a second residual processing layer, wherein the first residual processing layer and the second residual processing layer are both composed of a first residual convolution layer, a second residual convolution layer and a third residual convolution layer, the first residual convolution layer, the second residual convolution layer and the third residual convolution layer are sequentially connected, and the input feature map of the first residual convolution layer is also added to the output feature map of the third residual convolution layer by a jump connection to form an output feature map of the residual processing layer, and the residual processing layer performs feature extraction and dimensionality reduction processing on the feature map of the input residual module; The output of the first residual processing layer is used as the input of the second residual processing layer, wherein the specific calculation expression of each residual processing layer is as follows: y1=ReLU(BN(Conv 1×1,p (x))) Among them, Conv 1×1,p Indicates that a convolution process with a step size of p is performed on the input x of the residual processing layer using a convolution kernel of size 1×1, wherein the step size p is 1 by default; in the second feature extraction of the shallow residual module, the middle residual module and the deep residual module, the step size p in the convolution kernel of the first residual convolution layer in the first residual processing layer is set to 2; The calculation expression of the second residual convolution layer is: y2=ReLU(BN(Conv i×i,1 (y1))) Among them, Conv i×i Indicates that a convolution process with a step size of 1 is performed on the feature map y1 output by the first residual convolution layer using a convolution kernel of size i×i, wherein the size of the convolution kernel is 3×3 by default, but in the shallow residual module, the size of the convolution kernel is 5×5; The calculation expression of the third residual convolution layer is: <h2 style=";text-align:left;direction:ltr">y3=BN(Conv<h2 style=";text-align:left;direction:ltr"> 1×1,1 <h2 style=";text-align:left;direction:ltr"> (y2)) Among them, Conv 1×1 Indicates that a convolution process with a step size of 1 is performed on the feature map y2 output by the second residual convolution layer using a convolution kernel with a size of 1×1, so as to extract features from the feature map y2 output by the second residual convolution layer; The input feature map x of the first residual convolution layer is added to the output feature map y3 of the third residual convolution layer by skip connection to form the output feature map y4 of the residual processing layer. The calculation expression is: x skip =Conv 1×1,p (x) x skip =BN(x skip ) y4=ReLU(y3+x skip ) Among them, Conv 1×1,p It means that a convolution with a step size of p is performed on the input x of the residual processing layer using a convolution kernel of size 1×1, where the step size p is 1 by default, but the step size p in the second feature extraction of the shallow residual module, the middle residual module and the deep residual module is 2, that is, the input x is subjected to dimensionality reduction processing so that the dimension of the input feature map x matches the dimension of the output y3 of the third residual convolution layer so that they can be added smoothly.

8. The neural network model for engine cylinder liner character recognition according to claim 7 is characterized in that: The feature map output by the shallow residual module is reduced to the same dimension as the feature map output by the middle residual module and concatenated with the feature map output by the middle residual module and then input into the deep residual module.

9. A device for character recognition on an engine cylinder liner, characterized in that: Includes cameras, neural network models, and assembly equipment; The camera is used to capture the image on the engine cylinder liner to obtain the engine cylinder liner image with character spray code; The assembly equipment is used to install the engine cylinder liner on the engine; The neural network model is a neural network model according to any one of claims 5 to 8; the engine cylinder liner image with character coding is input into the neural network model for feature extraction and recognition classification, and a character string corresponding to the character coding on the engine cylinder liner image with character coding is output, and the character string is used to determine whether the engine cylinder liner matches the engine; If they match, a signal is transmitted to the assembly device to install the engine cylinder liner on the engine; if they do not match, the engine cylinder liner is not installed on the engine.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method for recognizing characters on an engine cylinder liner according to any one of claims 1 to 4 is implemented.

Images