A deep learning-based digital recognition method and system

Abstract

The application discloses a kind of based on deep learning's digital identification method and system, method includes image acquisition, image optimization model design, digital number dimension flexible unit design, digital bit field navigation unit design, digital image recognition model establishment and digital identification.The application belongs to the field of image processing, specifically refers to a kind of based on deep learning's digital identification method and system, the scheme is through digital number dimension flexible unit, according to the dynamic adjustment receptive field of digital real size, based on digital bit field navigation unit, through forward control and reverse refinement, improve the accuracy of digital positioning;Based on contour weighted loss and morphological perception loss, strengthen the segmentation integrity of digital contour fuzzy area;Through double-way synchronous optimization, balance key stroke detail preservation and anti-interference ability;Based on the weight generated by digital pixel peripheral zone dispersity, and introduce structure difference loss, significantly improve the feature learning of digital contour area;Further improve digital identification effect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a digital recognition method and system based on deep learning. BACKGROUND

[0002] The digital recognition method is to obtain digital related images or data, pre-process, feature extraction, and then use model matching or classification algorithm to determine the digital category. However, the general digital recognition method has the problems of digital local detail loss, incomplete structure capture, interference by non-digital contour, low overall positioning accuracy, and poor recognition accuracy. The general digital recognition method has the problems of poor anti-interference ability, loss of pixel-level details, insufficient distinction of key regions and similar digits, and poor digital recognition effect. SUMMARY

[0003] In view of the above problems, in order to overcome the defects of the prior art, the present application provides a digital recognition method and system based on deep learning. For the problem of digital local detail loss, incomplete structure capture, interference by non-digital contour, low overall positioning accuracy, and poor recognition accuracy of the general digital recognition method, the present application dynamically adjusts the receptive field according to the real size of the digit through the digital number dimension flexible matching unit, improves the digital feature extraction effect; based on the digital bit domain navigation unit, through forward guidance and control and reverse improvement, reduces background interference and regional misjudgment, improves the accuracy of digital positioning; based on contour weighted loss and morphological perception loss, strengthens the segmentation integrity of the fuzzy area of the digital contour, reduces the miss rate and misclassification rate; and further improves the subsequent digital recognition accuracy. For the problem of poor anti-interference ability, loss of pixel-level details, insufficient distinction of key regions and similar digits, and poor digital recognition effect of the general digital recognition method, the present application balances the key stroke detail retention and anti-interference ability through double-path synchronous optimization; based on the generation of weight of digital pixel peripheral zone dispersion, and introducing structure difference loss, significantly improves the feature learning of the digital contour area, effectively distinguishes the similar digits through key structure constraint, reduces the misjudgment rate of similar digits; and further improves the digital recognition effect.

[0004] The technical scheme adopted by the present application is as follows: the present application provides a digital recognition method based on deep learning, which comprises the following steps:

[0005] Step S1: image acquisition;

[0006] Step S2: image optimization model design;

[0007] Step S3: design of digital number dimension flexible matching unit;

[0008] Step S4: design of digital bit domain navigation unit;

[0009] Step S5: digital image recognition model establishment;

[0010] Step S6: digital recognition.

[0011] Further, in step S1, the image collection is to obtain historical digital original images; and preprocessing is performed, including format conversion, size standardization and annotation preparation; and an initial digital image set is obtained.

[0012] Further, in step S2, the image optimization model design is to optimize the collected initial digital image set, and the image optimization model adopts a single backbone network combined with a double branch line structure, including: a backbone network for extracting basic features of the digital image; a segmentation branch line for outputting a segmentation probability map of a digital region based on the backbone network features; a contour detection branch line for outputting a probability map of a digital contour based on the backbone network features; a digital dimension adaptation unit connected with the backbone network and the double branch line; and a digital bit domain navigation unit connected with the segmentation branch line and the contour detection branch line.

[0013] Further, in step S3, the digital dimension adaptation unit design specifically includes:

[0014] Step S31: double receptive field feature extraction; twice convolution is performed on the input digital feature map: small convolution kernel operation to obtain small receptive field feature F s ; large convolution kernel and hole convolution operation to obtain large receptive field feature F l , capturing the global structure of the digital;

[0015] Step S32: weight generation; F s and F l are spliced along the channel to obtain F cat ; and global average pooling is performed to generate a channel statistics vector S; S is converted into an intermediate vector Z through a fully connected layer, and then mapped into the weights W s of F s and the weights W l of F l through two independent fully connected layers, and normalized through Softmax; the normalized weights are multiplied with the original features element by element to obtain the features of the adapted digital scale.

[0016] Further, in step S4, the digital bit domain navigation unit design specifically includes:

[0017] Step S41: digital positioning mask generation; digital features F seg are obtained from the segmentation branch line; channel attention is used for screening to obtain ; spatial attention is used to position the digital position: the mean and maximum feature maps of the channel dimension are calculated for to generate a spatial mask , obtaining ; to thresholding and resizing, generating a digital positioning mask ;

[0018] Step S42: dual-track guidance architecture design; forward guidance: the and contour features element-wise multiplication, obtaining the features of the focused digital contour ; reverse refinement: when the loss of the contour detection branch is backpropagated, the gradient is fed back to the segmentation feature F seg , and the segmentation branch gradient is fused; the initial contour feature is the segmentation probability map of the digital region;

[0019] Step S43: image optimization loss function design; design contour weighted loss and morphology perception loss; generate weight through real label of contour detection branch; construct contour weighted loss; dynamically adjust parameters through area of digital real region; and quantify morphology complexity through number of closed regions and total curvature of contour, to obtain morphology perception loss; and then construct total image optimization loss function.

[0020] Further, in step S5, the digital image recognition model specifically includes the following contents:

[0021] Step S51: model architecture design; input the optimized digital image, and after double-path synchronous optimization processing, purify the features, and then input to the encoding-decoding process of U-shaped architecture, including encoder and decoder; finally output the probability map of 10 classes of numbers through convolution; obtain the digital recognition result;

[0022] Step S52: double-path synchronous optimization is to decompose the digital image into low-frequency components and high-frequency components with a Gaussian pyramid; use a Gaussian bandpass filter to retain details; expand the high-frequency components through Taylor expansion; the anti-interference path generates a morphing image from the digital original image, calculates the morphological change vector of the original image and the morphing image; converts the optical flow field into a 64-dimensional vector; splices the features of the detail path and the features of the anti-interference path into fused features;

[0023] Step S53: feature purification; replace the pooling layer with a convolution with a step size of 2, output after convolution of the fused features; and filter the features from two dimensions of channels and space; channel attention to the input features, compress the spatial dimension through global average pooling and maximum pooling, generate channel weights through MLP, and the weighted features are represented as F'; spatial attention to F' does average pooling and maximum pooling in the channel dimension, splices and generates spatial weights through convolution; obtain the weighted features;

[0024] Step S54: loss function design; a dynamic weighted loss function based on stroke contour is constructed, the dispersion of each pixel and its upper, lower, left and right surrounding area is calculated, an exponential normalized weight is established, a structure difference loss term is introduced, and a final dynamic weighted loss function is constructed.

[0025] Further, in step S6, the digital recognition is based on real-time digital image recognition by the image optimization model and the digital image recognition model.

[0026] The application provides a digital recognition system based on deep learning, which comprises an image acquisition module, an image optimization model design module, a digital number dimension adaptive unit design module, a digital bit domain navigation unit design module, a digital image recognition model establishment module and a digital recognition module.

[0027] The image acquisition module acquires historical digital original images and constructs an initial digital image set.

[0028] The image optimization model design module designs an image optimization model, which comprises a digital number dimension adaptive unit and a digital bit domain navigation unit.

[0029] The digital number dimension adaptive unit design module extracts double receptive field features and constructs features adaptive to digital scales.

[0030] The digital bit domain navigation unit design module generates a digital positioning mask, realizes double-track guidance and control through forward guidance and control and reverse improvement, and constructs an image optimization loss function.

[0031] The digital image recognition model establishment module establishes a digital image recognition model through double-path synchronous optimization and dynamic weighted loss design based on stroke contour.

[0032] The digital recognition module recognizes real-time digital images.

[0033] The application has the following beneficial effects by adopting the above scheme:

[0034] (1) In view of the problems of general digital recognition methods, such as loss of digital local details, incomplete structure capture, interference of non-digital contours, low overall positioning accuracy and poor recognition accuracy, the digital number dimension adaptive unit dynamically adjusts the receptive field according to the real size of the digital, improves the digital feature extraction effect, the digital bit domain navigation unit reduces background interference and regional misjudgment through forward guidance and control and reverse improvement, improves the accuracy of digital positioning, the contour weighted loss and the morphological perception loss strengthen the segmentation integrity of the fuzzy area of the digital contour, reduce the missed detection rate and the misclassification rate, and further improve the subsequent digital recognition accuracy.

[0035] (2) In view of the problems that the general digital recognition method has poor anti-interference ability, loses pixel-level details, and is insufficient in distinguishing key regions and similar numbers, thereby leading to poor digital recognition effect, the scheme balances key stroke detail retention and anti-interference ability through double-path synchronous optimization; the weight is generated based on the dispersion degree of the number pixel peripheral area, and the structural difference loss is introduced, which significantly improves the feature learning of the number contour area, effectively distinguishes similar numbers through key structure constraint, and reduces the misjudgment rate of similar numbers; and then improves the digital recognition effect. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 A flowchart of a digital recognition method based on deep learning provided by the present application is shown in the figure.

[0037] Figure 2 A schematic diagram of a digital recognition system based on deep learning provided by the present application is shown in the figure.

[0038] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, and are used to explain the present application together with embodiments of the present application, and do not constitute a limitation on the present application. DETAILED DESCRIPTION

[0039] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0040] In the description of the present application, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inner", "outer", and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the systems or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0041] Embodiment one, refer to Figure 1 The present application provides a digital recognition method based on deep learning, which comprises the following steps:

[0042] Step S1: image acquisition; acquire historical digital original images, and construct an initial digital image set;

[0043] Step S2: image optimization model design; design an image optimization model, including a digital number dimension matching unit and a digital bit domain navigation unit;

[0044] Step S3: Digital number dimension adaptive unit design; extract double receptive field features and construct adaptive digital scale features;

[0045] Step S4: Digital bit domain navigation unit design; generate digital positioning mask, realize double-track guidance through forward guidance and reverse refinement, and construct image optimization loss function;

[0046] Step S5: Digital image recognition model establishment; through double-path synchronous optimization and dynamic weighted loss design based on stroke contour, establish digital image recognition model;

[0047] Step S6: Digital recognition; recognize real-time digital images.

[0048] Example two, see Figure 1 This example is based on the above example, in step S1, image acquisition is to obtain historical digital original images, covering typical scenarios of digital recognition, including printed body, handwritten body and complex background; and preprocessing, including format conversion (uniformly converting the collected images to PNG format), size standardization and annotation preparation, including segmentation label (1 of binary graph represents digital area, 0 represents background) and contour label (1 of binary graph represents digital contour, 0 represents non-contour); obtain initial digital image set.

[0049] Example three, see Figure 1 This example is based on the above example, in step S2, image optimization model design is to optimize the collected initial digital image set, the image optimization model adopts single backbone network combined with double branch line structure, including: backbone network, used to extract the basic features of digital image, using ResNet; segmentation branch line, based on the feature output of backbone network, output the segmentation probability graph of digital area, to distinguish digital and background; contour detection branch line: based on the feature output of backbone network, output the probability graph of digital contour, to distinguish digital contour and non-contour; digital number dimension adaptive unit, connecting backbone network and double branch line, dynamically adjusting receptive field; digital bit domain navigation unit, connecting segmentation branch line and contour detection branch line, realizing double-track feature and gradient guidance.

[0050] Example four, see Figure 1 This example is based on the above example, in step S3, digital number dimension adaptive unit design is to dynamically adjust the receptive field according to the real size of the number, so that the receptive field range is accurately matched with the height-width ratio and pixel scale of the number, so as to capture the complete contour and internal structure of numbers of different sizes; the specific operation is:

[0051] Step S31: Double receptive field feature extraction; two convolutions are performed on the input digital feature map: small convolution kernel (3x3) operation, to obtain small receptive field feature F sFocusing on local details of the data; using large convolutional kernels (7×7) and dilated convolution (dilation=2) operations to obtain large receptive field features F. l Capture the global structure of the numbers;

[0052] Step S32: Weight generation; F s With F l Splicing along the channel into F cat Global average pooling is then performed to generate a channel statistics vector S, representing global information about different receptive field features. S is then transformed into an intermediate vector Z via a fully connected layer, and finally mapped to F via two independent fully connected layers. s weight W s and F l weight W l The features are then normalized using Softmax; the normalized weights are multiplied element-wise with the original features to obtain features adapted to the digital scale.

[0053] Example 5, see Figure 1 This embodiment is based on the above embodiment. In step S4, the digital bit-domain navigation unit design uses segmentation features to locate the digital region and suppress non-digital contours; it uses contour features to correct segmentation errors and improve digital positioning accuracy. The specific operation is as follows:

[0054] Step S41: Digital localization mask generation; obtaining digital features F from segmentation branches. seg This includes information on the regional distribution of numbers; channel attention is used to filter F. seg By learning channel weights, we strengthen channels related to numbers and weaken background channels, thus obtaining... Using spatial attention to locate numbers: Calculate the mean and maximum feature maps along the channel dimension to generate a spatial mask. Highlighting the spatial region where the numbers are located, we obtain ;in, It is the segmentation feature map after channel filtering; It is a channel attention architecture; It is a segmentation feature map after spatial localization; for Perform thresholding (preserving regions greater than 0.5) and resize to generate a digital localization mask with the same size as the contour feature map. ;

[0055] Step S42: Dual-track control architecture design; Forward control: ... Contour features Element-wise multiplication suppresses background interference and similar digit contours, resulting in the feature of the focused digit contour. , represented as: ; Backward refinement: when the loss of the contour detection branch is backpropagated, the gradient is fed back to the segmentation feature F seg , which is fused with the gradient of the segmentation branch to correct the segmentation error; the gradient optimization of the segmentation feature is expressed as: ; wherein, is the updated gradient of the segmentation feature; is the gradient of the segmentation branch loss; is the gradient of the contour detection branch loss; is the gradient balance coefficient; the initial contour feature is the segmentation probability map of the digital region;

[0056] Step S43: image optimization loss function design; for the problems of small target digital missed detection, incomplete segmentation caused by fuzzy contour, and similar digital missegmentation in the digital recognition scene, contour weighted loss and morphological perception loss are designed; the contour detection branch weight is generated by the real label, expressed as: ; contour weighted loss is expressed as: ; the dynamic parameter is adjusted by the area of the digital real region, expressed as: ; ; wherein, is the contour loss pixel weight, is the pixel coordinate on the feature map; is the contour weight coefficient; is the contour real label; is the pixel-level real label of the segmentation branch; H and W are height and width; is the standard focus loss; is the digital region probability predicted by the segmentation branch; is the missed judgment penalty coefficient; is the misjudgment penalty coefficient; is the gain coefficient; is the area of the digital real region, taking the number of pixels; T is the area threshold; and the morphological complexity is quantified by the number of closed regions and the total curvature of the contour, and the loss of complex numbers is amplified, expressed as: ; is the complexity weight coefficient; X is the normalized intersection point number, the total number of pixel points of the intersection of the strokes in the digital contour contour, and the intersection point number is obtained by calculating the self-intersection point of the contour line; B is the normalized branch number after skeletonization of the digital skeleton, and the branch number is obtained by skeleton endpoint detection; morphological perception loss is expressed as: ; total image optimization loss function is expressed as: ; wherein, t and p are the real label vector and the predicted probability vector of the contour detection, respectively; is the smoothing term.

[0057] By performing the above operation, for the general digital recognition method, there are problems of loss of local details of numbers, incomplete structure capture, interference by non-number contours, low overall positioning accuracy, and further poor recognition accuracy. The scheme dynamically adjusts the receptive field according to the real size of the number through the number dimension matching unit, improves the number feature extraction effect; based on the number bit field navigation unit, through forward guidance and control and reverse improvement, reduces background interference and regional misjudgment, improves the accuracy of number positioning; based on the contour weighted loss and the morphological perception loss, the segmentation integrity of the fuzzy area of the number contour is strengthened, and the miss rate and the misclassification rate are reduced; and then the subsequent number recognition accuracy is improved.

[0058] Embodiment six, refer to Figure 1 This embodiment is based on the above embodiment. In step S5, the digital image recognition model specifically includes the following contents:

[0059] Step S51: model architecture design; input the optimized digital image, perform feature purification after double-path synchronous optimization processing, then input to the encoding-decoding process of U-shaped architecture, including encoder (core feature extraction) and decoder (enhanced feature extraction); the encoder is combined by 3x3 convolution (padding=1, step=1) and 3x3 convolution with step=2, and generates 5 feature layers through 4 times of dimension reduction to capture features from local strokes to overall morphology; the decoder enlarges the high-level features through up-convolution, and connects with the same size feature layers of the encoder through jump connection, and then enhances the contour features through 2 times of 3x3 convolution, and gradually recovers to 28x28 size; finally, 1x1 convolution is used to output the probability map of 10 types of numbers (0-9); and the digital recognition result is obtained;

[0060] Step S52: double-path synchronous optimization is to enhance the fine features of strokes through the detail path, and extract the anti-interference morphological features through the anti-interference path, and finally realize the complementation of detail-robust information; the detail path decomposes the digital image into low-frequency components (overall morphology) and high-frequency components (details) by using the Gaussian pyramid; the effective details are reserved by using the Gaussian band-pass filter; the high-frequency components are expanded by using Taylor expansion to avoid excessive amplification of noise, and are expressed as: The anti-interference path generates a deformation image (including translation of 1 pixel, rotation of 5°, and scaling of 0.95 times) from the original digital image, calculates the morphological change vector of the original image and the deformation image by using the Farneback optical flow algorithm, and the energy function E is expressed as: The optical flow field is converted into a 64-dimensional vector; the 64-dimensional features (stroke details) of the detail path and the 64-dimensional features (morphological robustness) of the anti-interference path are spliced into 128-dimensional fusion features; wherein, is the gray value after detail enhancement; is the original gray value (optimized digital image); is the detail amplification factor; is the gradient; is the spatial domain of the digital image (28x28 pixel range); is the gray matrix of the optimized digital image; is the deformed digital image gray matrix; U is the optical flow field vector;

[0061] Step S53: feature purification; the digital subtle features are the key to classification, to avoid information loss, background noise introduces redundant features, replace the pooling layer with convolution step = 2, while reducing dimension, retain pixel-level details, for 128-dimensional fusion features, after 3x3 convolution (step = 2, padding = 1) the output size is ; and filter features from two dimensions of channel and space, channel attention to input features, through global average pooling and maximum pooling, compress the spatial dimension, generate channel weight through MLP, the weighted feature is represented as F'; spatial attention to F' in the channel dimension does average pooling and maximum pooling, after splicing, generate spatial weight through 3x3 convolution, the weighted feature is represented as F'';

[0062] Step S54: loss function design; construct a dynamic weighted loss function based on stroke contour, the pixel value of digital stroke contour changes most sharply, so for the digital true segmentation label, calculate the dispersion of each pixel and its upper, lower, left and right neighborhood, quantify the contour intensity, represented as: ; Establish an exponential normalization weight, the more intense the contour, the greater the weight, represented as: ; And introduce a structural difference loss term , quantify the prediction difference of similar numbers on key structures, calculate additional loss, avoid misjudgment caused by similar numbers due to key structure confusion, the final dynamic weighted loss function is represented as: ; ; Wherein, is the stroke contour dispersion of pixel ; is the true segmentation label of pixel ; is the average value of all pixel true labels in the neighborhood; N is the number of pixels in the neighborhood; is the pixel dynamic weight; is the class balance coefficient, when , ; When , ; is the probability that the predicted pixel belongs to the digital area; is the focus parameter; is a key structure area binary mask of category m, including the value 1 of the number 0 in the bottom 1 / 4 arc area; the value 1 of the number 1 in the upper 1 / 2 vertical line area; the value 1 of the number 2 in the right end point and the bottom connected pen area; the value 1 of the number 3 in the left end point and the upper arc area; the value 1 of the number 4 in the intersection point and the right side 1 / 2 vertical line area; the value 1 of the number 5 in the right upper corner horizontal broken corner and the bottom horizontal line area; the value 1 of the number 6 in the bottom 1 / 3 curve area; the value 1 of the number 7 in the upper horizontal pen area; the value 1 of the number 8 in the middle 1 / 5 horizontal area; the value 1 of the number 9 in the top 1 / 3 curve area; and the value 0 of other areas; is the gradient direction of the prediction probability map; is the gradient direction of the real segmentation label;

[0063] Regarding model training: for a set of digital images, the images are divided into a training set, a validation set and a test set, the image optimization model is divided into the initial set of digital images; the digital image recognition model is divided into the optimized initial set of digital images; the training set is used for model parameter learning, the validation set is used for monitoring overfitting in the training process, and the hyperparameters are adjusted; the test set is independent of the training process and is used for final evaluation of the generalization ability of the model; the Adam optimizer is used for parameter updating; and the overall accuracy and the contour area accuracy are used to verify the performance of the model.

[0064] By performing the above operations, the problems of poor anti-interference ability, loss of pixel-level details, insufficient distinction between key regions and similar numbers, and poor digital recognition effect of general digital recognition methods are solved. The present scheme balances the preservation of key stroke details and the anti-interference ability through double-path synchronous optimization; generates weights based on the pixel peripheral zone dispersion of the number, and introduces a structural difference loss, which significantly improves the feature learning of the number contour area, effectively distinguishes similar numbers through key structure constraints, and reduces the misjudgment rate of similar numbers; and further improves the digital recognition effect.

[0065] Embodiment seven, refer to Figure 1 This embodiment is based on the above-mentioned embodiment, in step S6, the digital recognition is based on the image optimization model and the digital image recognition model to recognize the real-time digital image; after the real-time digital image is preprocessed, it is input into the image optimization model and the digital image recognition model in turn, and the output of the digital image recognition model is taken as the digital recognition result.

[0066] Embodiment eight, refer to Figure 2 This embodiment is based on the above-mentioned embodiment, and the present application provides a digital recognition system based on deep learning, which comprises an image acquisition module, an image optimization model design module, a digital number dimension matching unit design module, a digital bit domain navigation unit design module, a digital image recognition model establishment module and a digital recognition module.

[0067] The image acquisition module acquires historical digital original images, and constructs an initial digital image set;

[0068] The image optimization model design module designs an image optimization model, including a digital dimension adaptation unit and a digital bit domain navigation unit;

[0069] The digital dimension adaptation unit design module extracts double-receptive field features, and constructs features of adaptive digital scales;

[0070] The digital bit domain navigation unit design module generates a digital positioning mask, realizes double-track guidance and control through forward guidance and control and reverse refinement, and constructs an image optimization loss function;

[0071] The digital image recognition model establishment module establishes a digital image recognition model through double-path synchronous optimization and dynamic weighted loss design based on stroke contours;

[0072] The digital recognition module recognizes real-time digital images.

[0073] It should be noted that in this document, the terms "comprising", "including", or any other variant thereof are intended to cover non-exclusive inclusions, so that processes, methods, articles, or devices that include a series of elements not only include those elements, but also include other elements not explicitly listed, or inherent to such processes, methods, articles, or devices.

[0074] Although embodiments of the present application have been shown and described, it will be understood by those having ordinary skill in the art that various changes, modifications, substitutions and alterations can be made hereto without departing from the principles and spirit of the present application.

[0075] The above describes the present application and its embodiments, which are not restrictive, and the embodiments shown in the drawings are only one of the embodiments of the present application, and the actual structure is not limited thereto. In general, if a person skilled in the art is inspired by it, without departing from the purpose of the present application, without creative design, similar structure and embodiments of the technical solution can be designed, which should belong to the protection scope of the present application.

Claims

1. A deep learning-based digital recognition method, characterized in that: The method comprises the following steps: Step S1: image acquisition; acquire historical digital original images, and construct an initial digital image set; Step S2: image optimization model design; design an image optimization model, including a digital dimension adaptation unit and a digital bit domain navigation unit; Step S3: design of the digital dimension adaptation unit; extract double-receptive-field features, and construct features adapted to digital scales; Step S4: design of the digital bit domain navigation unit; generate a digital positioning mask, realize double-track guidance through forward guidance and reverse refinement, and construct an image optimization loss function; Step S5: digital image recognition model establishment; establish a digital image recognition model through double-path synchronous optimization and dynamic weighted loss design based on stroke contours; Step S6: digital recognition; recognize real-time digital images; In step S4, the design of the digital bit domain navigation unit specifically comprises: Step S41: digital positioning mask generation; obtain digital features F from the segmentation branch seg ; filtered by channel attention, to obtain ; position the digital position with spatial attention: to Calculate the mean and maximum feature map of the channel dimension to generate a spatial mask , to obtain ; threshold processing is performed on , and the size is adjusted to generate a digital positioning mask ; wherein, is the segmentation feature map after channel filtering; is the segmentation feature map after spatial positioning; Step S42: dual-track guided architecture design; forward guidance: the contour feature is obtained by multiplying the segmentation probability map and the contour feature with the contour feature Element-wise multiplication to get the feature of the focused digital contour ; backward refinement: when the loss of the contour detection branch is backpropagated, the gradient is fed back to the segmentation feature F seg , and the gradient of the segmentation branch is fused; the initial contour feature is the segmentation probability map of the digital region Step S43: image optimization loss function design; design contour weighting loss and morphology perception loss; generate weights through real labels of contour detection branch lines; construct contour weighting loss; dynamically adjust parameters through the area of digital real regions; and quantify the morphology complexity through the number of closed regions and the total curvature of contours to obtain the morphology perception loss; and further construct a total image optimization loss function; In step S5, the establishment of the digital image recognition model specifically comprises the following contents: Step S51: model architecture design; input the optimized digital image, purify features after double-path synchronous optimization, then input to the encoding-decoding process of the U-shaped architecture, including an encoder and a decoder; finally output a probability map of 10 classes of numbers through convolution; and obtain a digital recognition result; Step S52: double-path synchronous optimization is to decompose the digital image into low-frequency components and high-frequency components through a Gaussian pyramid; retain details through a Gaussian bandpass filter; expand the high-frequency components through Taylor expansion; generate a morphing image for the digital original image through the disturbance path; calculate the morphological change vector of the digital original image and the morphing image through the Farneback optical flow algorithm; convert the optical flow field into a 64-dimensional vector; and splice the features of the detail path and the features of the disturbance path into fusion features; Step S53: feature purification; replace the pooling layer with a convolution with a step size of 2, output the convolution of the fusion features; and filter features from two dimensions of channels and space; the channel attention calculates the input features, compresses the spatial dimension through global average pooling and maximum pooling, generates channel weights through an MLP, and represents the weighted features as F'; the spatial attention calculates the average pooling and maximum pooling of F' in the channel dimension, splices them, and generates spatial weights through convolution; and obtain the weighted features; Step S54: loss function design; construct a dynamic weighted loss function based on stroke contours, calculate the dispersion of each pixel and its upper, lower, left and right neighborhood, establish an exponential normalization weight, and introduce a structure difference loss term to construct a final dynamic weighted loss function. 2.The deep learning-based digital recognition method of claim 1, wherein: In step S2, the image optimization model design is to optimize the initial digital image set collected, and the image optimization model adopts a single backbone network combined with a double branch line structure, including: a backbone network for extracting basic features of the digital image; a segmentation branch line for outputting a segmentation probability map of the digital region based on the backbone network feature; a contour detection branch line for outputting a probability map of the digital contour based on the backbone network feature; a digital number dimension adaptive unit connected with the backbone network and the double branch line; and a digital bit domain navigation unit connected with the segmentation branch line and the contour detection branch line. 3.The deep learning-based digital recognition method of claim 2, wherein: In step S3, the digital number dimension adaptive unit specifically includes: Step S31: double receptive field feature extraction; twice convolution is performed on the input digital feature map: small convolution kernel operation to obtain small receptive field feature F s ; large convolution kernel and cavity convolution operation to obtain large receptive field feature F l , capturing the global structure of the digital; Step S32: Weight generation; F s With F l Splicing along the channel into F cat Global average pooling is then performed to generate a channel statistics vector S; S is transformed into an intermediate vector Z through a fully connected layer, and then mapped to F through two independent fully connected layers. s weight W s and F l weight W l The features are then normalized using Softmax; the normalized weights are multiplied element-wise with the original features to obtain features adapted to the digital scale.

4. The deep learning-based digital recognition method of claim 3, wherein: In step S1, the image acquisition is to obtain historical digital original images, and to perform preprocessing including format conversion, size standardization and label preparation, thereby obtaining an initial digital image set.

5. The deep learning-based digital recognition method of claim 4, wherein: In step S6, the digital recognition is to recognize real-time digital images based on the image optimization model and the digital image recognition model.

6. A deep learning based digital recognition system for implementing a deep learning based digital recognition method according to any one of claims 1-5, characterized in that: The system includes an image acquisition module, an image optimization model design module, a digital number dimension adaptive unit design module, a digital bit domain navigation unit design module, a digital image recognition model establishment module and a digital recognition module. The image acquisition module acquires historical digital original images and constructs an initial digital image set. The image optimization model design module designs an image optimization model including a digital number dimension adaptive unit and a digital bit domain navigation unit. The digital number dimension adaptive unit design module extracts double receptive field features and constructs features adaptive to digital scales. The digital bit domain navigation unit design module generates a digital positioning mask, realizes double-track guidance and control through forward guidance and control and reverse refinement, and constructs an image optimization loss function. The digital image recognition model establishment module establishes a digital image recognition model through double-path synchronous optimization and dynamic weighted loss design based on stroke contours. The digital recognition module recognizes real-time digital images.

Images