A deep learning-based image recognition method and system
By calculating image preprocessing coefficients and dynamically selecting preprocessing strategies, different preprocessing methods are adopted for images of different qualities. Combined with deep learning models, this solves the problem of poor efficiency and effectiveness of image preprocessing in existing technologies, and achieves efficient image recognition.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- LUDONG UNIVERSITY
- Filing Date
- 2024-06-28
- Publication Date
- 2026-07-24
AI Technical Summary
In existing technologies, image preprocessing algorithms need to be specified in advance and cannot be selected according to the image conditions, resulting in poor image processing effects for low-quality images and low efficiency for high-quality image processing.
By calculating the preprocessing coefficients of the image, a suitable preprocessing strategy is dynamically selected. Two preprocessing algorithms are used for images with poor quality, while only one algorithm is used for images with high quality. The images are then combined with a deep learning model for recognition.
It improves the efficiency and effectiveness of image preprocessing, ensuring that images of poor quality are effectively improved, while avoiding unnecessary processing of high-quality images, thus improving the overall recognition efficiency.
Smart Images
Figure CN118736381B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image recognition, and more particularly to an image recognition method and system based on deep learning. Background Technology
[0002] In existing technologies, image preprocessing is typically required to improve image quality before deep learning can be used for recognition. Image preprocessing can involve using only one algorithm or two separate algorithms, followed by merging the results to obtain the preprocessed image. Inputting this preprocessed image into a deep learning model then yields the recognition result.
[0003] However, existing technologies have the following drawbacks: the preprocessing algorithms used need to be specified in advance and cannot be selected based on the image conditions. This makes it more likely that only one preprocessing algorithm will be used for images of poor quality, resulting in limited improvement in image quality. At the same time, multiple preprocessing algorithms may be used for images of good quality, which will affect the efficiency of preprocessing and thus the overall efficiency of image recognition. Summary of the Invention
[0004] The purpose of this invention is to disclose an image recognition method and system based on deep learning, which solves the problem of how to select a preprocessing algorithm based on the actual situation of the image, thereby improving the overall image preprocessing efficiency while effectively improving the quality of poor-quality images.
[0005] To achieve the above objectives, the present invention provides the following technical solution:
[0006] In a first aspect, the present invention provides an image recognition method based on deep learning, comprising:
[0007] S1, acquire the image to be recognized;
[0008] S2, calculate the preprocessing coefficients of the image to be recognized;
[0009] S3, determine the preprocessing strategy for the image to be recognized based on the preprocessing coefficients;
[0010] S4, based on the preprocessing strategy, preprocess the image to be recognized to obtain the preprocessed image;
[0011] S5 inputs the preprocessed image into a deep learning-based image recognition model for recognition.
[0012] Preferably, the calculation process of the preprocessing coefficients includes:
[0013] The first step is to convert the image to grayscale to obtain a grayscale image.
[0014] The second step is to obtain the computation region in the grayscale image;
[0015] The third step is to calculate the preprocessing coefficients of the image to be identified based on the computational region.
[0016] Preferably, the image to be recognized is converted to grayscale to obtain a grayscale image, including:
[0017] The image to be identified is converted to grayscale using a weighted average method to obtain a grayscale image.
[0018] Preferably, the image to be identified is converted to grayscale using a weighted average method to obtain a grayscale image, including:
[0019] The following formula is used to calculate the grayscale image for each pixel in the image to be recognized:
[0020] gray b =w1×R b +w2×G b +w3×B b
[0021] gray b R represents the grayscale value of pixel b. b G b and B b represents the red component, green component, and pixel value of the corresponding red component in the image of pixel b in the RGB color space, respectively, and w1, w2, and w3 are the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient, respectively.
[0022] Preferably, obtaining the computation region in the grayscale image includes:
[0023] The image to be identified is divided into N×N regions of equal area;
[0024] Based on the distance from the center of the image to be identified, from near to far, calculate the region judgment value for each region sequentially. Starting from the third region judgment value, perform the following processing:
[0025] Determine the value of the (k+1)th region: arejud k+1 Does arejud satisfy? k+1 >arejud k >arejud k-1 If so, then arejud k+1 The corresponding region is used as the calculation region, arejud kand arejud k-1 These represent the judgment values for the k-th and (k-1)-th regions, respectively; k is greater than or equal to 2.
[0026] If not, check if k is equal to N. 2 If not, continue calculating the region judgment value for the next region; if yes, take the region with the largest region judgment value as the calculation region.
[0027] Preferably, the formula for calculating the region judgment value is:
[0028]
[0029] arejud k This represents the judgment value of the k-th region obtained by calculating the k-th region, where Uk represents the set of pixels in the k-th region, nUk represents the total number of pixels in Uk, and gray i This represents the grayscale value of pixel i in Uk.
[0030] Preferably, the preprocessing coefficients of the image to be identified are calculated based on the computational region, including:
[0031] Calculate the preprocessing coefficients using the following formula:
[0032]
[0033] precoef represents the preprocessing coefficients, M1 represents the number of pixels in the computational region whose grayscale value is greater than the average grayscale value of the pixels in the computational region, M2 represents the number of pixels in image P obtained after processing the pixels in the computational region using the NLM algorithm whose grayscale value is greater than the average grayscale value of the pixels in P, ma1 represents the maximum grayscale value of the pixels in P, and m j M1 represents the total number of pixels with grayscale value j in P, M2 represents the total number of pixels in P, and m represents the maximum grayscale value in the calculation area. k M4 represents the number of pixels with a gray value of k in the calculation area, and δ1 and δ2 represent the first weight and the second weight, respectively.
[0034] Preferably, the preprocessing strategy for determining the image to be recognized based on the preprocessing coefficients includes:
[0035] If the preprocessing coefficient is greater than the preset coefficient threshold, the first preprocessing method is adopted as the preprocessing strategy for the image to be identified.
[0036] If the preprocessing coefficients are less than or equal to the preset coefficient threshold, then the second preprocessing method is used as the preprocessing strategy for the image to be identified.
[0037] Preferably, the first preprocessing method includes:
[0038] Let NP represent the image that needs to be preprocessed. For NP, the NLM algorithm and wavelet denoising algorithm are used to preprocess NP to obtain images NP1 and NP2 respectively.
[0039] The formula for obtaining the preprocessed image PP is as follows:
[0040] NP b =λ×NP1 b +(1-λ)×NP2 b
[0041] NP b NP1 represents the grayscale value of pixel b in image PP. b and NP2 b λ represents the grayscale value of the pixel in images NP1 and NP2 that has the same coordinate as pixel b in image PP, respectively, and λ represents the set acquisition coefficient, λ∈(0.2,0.8).
[0042] Secondly, the present invention provides an image recognition system based on deep learning, including an acquisition module, a calculation module, a determination module, a preprocessing module, and a recognition module;
[0043] The acquisition module is used to acquire the images that need to be recognized;
[0044] The calculation module is used to calculate the preprocessing coefficients of the image to be recognized;
[0045] The determination module is used to determine the preprocessing strategy for the image to be recognized based on the preprocessing coefficients;
[0046] The preprocessing module is used to preprocess the images to be recognized based on the preprocessing strategy to obtain preprocessed images;
[0047] The recognition module is used to input preprocessed images into a deep learning-based image recognition model for recognition.
[0048] Beneficial effects:
[0049] Compared to existing technologies, this invention, when preprocessing images input into a deep learning-based image recognition model, does not employ only one or two preprocessing algorithms. Instead, it selects an appropriate preprocessing algorithm based on the actual condition of the image. For images of poor quality, two preprocessing algorithms are used to ensure preprocessing effectiveness. When processing images of already high quality, it avoids using two different preprocessing algorithms, thus improving processing efficiency. Therefore, this invention achieves good preprocessing results while maintaining efficiency. Attached Figure Description
[0050] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0051] Figure 1 This is a schematic diagram of an image recognition method based on deep learning according to the present invention.
[0052] Figure 2 This is a schematic diagram of an image recognition system based on deep learning according to the present invention. Detailed Implementation
[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0054] like Figure 1 As shown, this invention provides an image recognition method based on deep learning, comprising:
[0055] S2, calculate the preprocessing coefficients of the image to be recognized;
[0056] S3, determine the preprocessing strategy for the image to be recognized based on the preprocessing coefficients;
[0057] S4, based on the preprocessing strategy, preprocess the image to be recognized to obtain the preprocessed image;
[0058] S5 inputs the preprocessed image into a deep learning-based image recognition model for recognition.
[0059] When preprocessing images input into a deep learning-based image recognition model, the aforementioned method does not employ only one or two preprocessing algorithms. Instead, it selects an appropriate preprocessing algorithm based on the actual image conditions. For images of poor quality, two preprocessing algorithms are used to ensure effective preprocessing. When processing images of already high quality, the use of two different preprocessing algorithms is avoided, thus improving processing efficiency. Therefore, this invention achieves good preprocessing results while maintaining efficiency.
[0060] Preferably, the calculation process of the preprocessing coefficients includes:
[0061] The first step is to convert the image to grayscale to obtain a grayscale image.
[0062] The second step is to obtain the computation region in the grayscale image;
[0063] The third step is to calculate the preprocessing coefficients of the image to be identified based on the computational region.
[0064] By obtaining the computation area, it is possible to avoid calculating preprocessing coefficients based on all pixels, thereby improving the efficiency of obtaining preprocessing coefficients.
[0065] Preferably, the image to be recognized is converted to grayscale to obtain a grayscale image, including:
[0066] The image to be identified is converted to grayscale using a weighted average method to obtain a grayscale image.
[0067] Image grayscale conversion is the process of converting a color image into a grayscale image. This is very useful for reducing data dimensionality, speeding up image processing, and meeting certain analytical needs. Besides the weighted average method, other methods such as the maximum value method and the average value method can also be used to convert images that need to be identified to grayscale.
[0068] Maximum value method: This method uses the largest value among the R, G, and B components of each pixel as its grayscale value. It is suitable for situations where you want to preserve details in the brightest parts of an image.
[0069] Average value method: Directly take the arithmetic mean of the R, G, and B components of the pixel as the gray value. This method is relatively simple.
[0070] Preferably, the image to be identified is converted to grayscale using a weighted average method to obtain a grayscale image, including:
[0071] The following formula is used to calculate the grayscale image for each pixel in the image to be recognized:
[0072] gray b =w1×R b +w2×G b +w3×B b
[0073] gray b R represents the grayscale value of pixel b. b G b and B b represents the red component, green component, and pixel value of the corresponding red component in the image of pixel b in the RGB color space, respectively, and w1, w2, and w3 are the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient, respectively.
[0074] Preferably, the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient are 0.32, 0.56, and 0.12, respectively.
[0075] Preferably, obtaining the computation region in the grayscale image includes:
[0076] The image to be identified is divided into N×N regions of equal area;
[0077] Based on the distance from the center of the image to be identified, from near to far, calculate the region judgment value for each region sequentially. Starting from the third region judgment value, perform the following processing:
[0078] Determine the value of the (k+1)th region: arejud k+1 Does arejud satisfy? k+1 >arejud k >arejud k-1 If so, then arejud k+1 The corresponding region is used as the calculation region, arejud k and arejud k-1 These represent the judgment values for the k-th and (k-1)-th regions, respectively; k is greater than or equal to 2.
[0079] If not, check if k is equal to N. 2 If not, continue calculating the region judgment value for the next region; if yes, take the region with the largest region judgment value as the calculation region.
[0080] In the process of obtaining the calculation region, the present invention aims to have a certain representative local area, that is, a region carrying certain image detail information. Therefore, the present invention divides the image into multiple regions of equal area, and then calculates the region judgment value based on the distance between the region and the center of the image, thereby avoiding the use of regions that do not contain image details as the calculation region, which is beneficial to improving the effectiveness of subsequent selection of preprocessing schemes.
[0081] Specifically, during the sorting process, if multiple regions are found to be equidistant from the center of the image, these regions are randomly sorted. For example, when different regions are sorted in ascending order of their distances from the center of the image, the distances would be 10, 15, 20, 25, and 30.
[0082] When there is only one region with a distance of 10, calculate the region judgment value for that region first. Then, if there are three regions with a distance of 15, randomly sort these three regions and calculate the region judgment value for these three sorted regions first. Then, calculate the region with a distance of 20. If there are multiple regions with a distance of 20, calculate the region coefficient according to the rules for calculating the region judgment value for the region with a distance of 15.
[0083] Preferably, the formula for calculating the region judgment value is:
[0084]
[0085] arejud k This represents the judgment value of the k-th region obtained by calculating the k-th region, where Uk represents the set of pixels in the k-th region, nUk represents the total number of pixels in Uk, and gray i This represents the grayscale value of pixel i in Uk.
[0086] The region judgment value of this invention is mainly represented by the fluctuation of gray values. The greater the fluctuation of gray values, the more image details are carried in the region, which is beneficial for selecting a representative calculation region.
[0087] Preferably, the preprocessing coefficients of the image to be identified are calculated based on the computational region, including:
[0088] Calculate the preprocessing coefficients using the following formula:
[0089]
[0090] precoef represents the preprocessing coefficients, M1 represents the number of pixels in the computational region whose grayscale value is greater than the average grayscale value of the pixels in the computational region, M2 represents the number of pixels in image P obtained after processing the pixels in the computational region using the NLM algorithm whose grayscale value is greater than the average grayscale value of the pixels in P, ma1 represents the maximum grayscale value of the pixels in P, and m j M1 represents the total number of pixels with grayscale value j in P, M2 represents the total number of pixels in P, and m represents the maximum grayscale value in the calculation area. k M4 represents the number of pixels with a gray value of k in the calculation area, and δ1 and δ2 represent the first weight and the second weight, respectively.
[0091] The preprocessing coefficients of this invention are calculated by comprehensively considering two factors: the number of pixels whose pixel values change significantly after processing the computational region using the NLM algorithm, and the changes in image details within the computational region before and after processing. This allows the preprocessing coefficients to more accurately represent the changes in the computational region before and after NLM processing. A larger preprocessing coefficient indicates a higher probability of effectively improving the quality of pixels in the computational region, meaning a lower quality original image to be recognized. Conversely, a smaller preprocessing coefficient indicates a lower probability of effectively improving pixel quality, meaning a higher quality original image to be recognized. Thus, the selection of subsequent preprocessing methods in this invention can be adaptively based on the preprocessing coefficients.
[0092] Preferably, the first weight is The second weight is
[0093] Preferably, the preprocessing strategy for determining the image to be recognized based on the preprocessing coefficients includes:
[0094] If the preprocessing coefficient is greater than the preset coefficient threshold, the first preprocessing method is adopted as the preprocessing strategy for the image to be identified.
[0095] If the preprocessing coefficients are less than or equal to the preset coefficient threshold, then the second preprocessing method is used as the preprocessing strategy for the image to be identified.
[0096] Preferably, the preset coefficient threshold is 0.1.
[0097] Preferably, the first preprocessing method includes:
[0098] Let NP represent the image that needs to be preprocessed. For NP, the NLM algorithm and wavelet denoising algorithm are used to preprocess NP to obtain images NP1 and NP2 respectively.
[0099] The formula for obtaining the preprocessed image PP is as follows:
[0100] NP b =λ×NP1 b +(1-λ)×NP2 b
[0101] NP b NP1 represents the grayscale value of pixel b in image PP. b and NP2 b λ represents the grayscale value of the pixel in images NP1 and NP2 that has the same coordinate as pixel b in image PP, respectively, and λ represents the set acquisition coefficient, λ∈(0.2,0.8).
[0102] By processing the noise reduction results from two different processing domains and then fusing the results, the quality of the preprocessed image can be significantly improved.
[0103] Wavelet denoising is a highly effective noise reduction technique in image processing, primarily based on the theory of wavelet transform. Wavelet transform decomposes an image into different frequency components, giving it localization properties in both time and frequency dimensions. This makes it ideal for analyzing and processing non-stationary signals, such as edge and texture features in images. Wavelet transform represents signals using a series of wavelet basis functions, obtained by translating and scaling a mother wavelet function. In image processing, two-dimensional wavelet transform is typically used, decomposing the image horizontally, vertically, and diagonally.
[0104] Wavelet noise reduction algorithm steps:
[0105] 1. Wavelet Decomposition: First, a wavelet transform is applied to the noisy image, decomposing it into different frequency subbands. This process is often called multi-resolution analysis because each level of decomposition produces a set of approximate (low-frequency) and detail (high-frequency) coefficients.
[0106] 2. Threshold processing: Since noise often appears in high-frequency detail coefficients, it can be suppressed by setting a threshold.
[0107] There are two main thresholding strategies: hard thresholding and soft thresholding.
[0108] Hard threshold: If the absolute value of a wavelet coefficient is less than the threshold, it is set to zero; otherwise, it remains unchanged.
[0109] Soft thresholding: If the absolute value of the wavelet coefficient is less than the threshold, it is set to zero; otherwise, the threshold is subtracted. This can smoothly remove noise and avoid the block effect that may be caused by hard thresholding.
[0110] 3. Wavelet reconstruction: After thresholding, the image is reassembled using inverse wavelet transform to obtain the denoised image.
[0111] Wavelet denoising algorithms can remove noise while preserving important features of the image such as edges and textures, thus achieving excellent results in both visual appeal and quality assessment.
[0112] Preferably, the set acquisition coefficient is 0.4.
[0113] Preferably, the second preprocessing method includes:
[0114] Let NP represent the image that needs to be preprocessed. We can directly use the NLM algorithm to preprocess NP to obtain the preprocessed image.
[0115] Preferably, the deep learning-based image recognition model is a convolutional neural network model.
[0116] Specifically, the convolutional neural network model of the present invention includes a base layer, a pooling layer, a fully connected layer, and an output layer.
[0117] Convolutional layers:
[0118] A convolutional kernel (also called a filter or feature detector) slides across an image, performing element-wise multiplication and summation to extract local features of the image, such as edges and textures.
[0119] Each convolutional kernel generates a new feature map, which represents the response intensity of the image in a specific direction or feature.
[0120] By using multiple convolutional kernels, the network can learn different types of features.
[0121] Pooling layer:
[0122] This is used to reduce the spatial dimension of feature maps, reduce computation, and maintain robustness to transformations and rotations.
[0123] The most common pooling method is max pooling, which takes the maximum value in the neighboring regions of the feature map as the output.
[0124] Activation function:
[0125] For example, the ReLU (Rectified Linear Unit) function is used to introduce non-linearity, enabling the network to learn more complex features.
[0126] Fully connected layer:
[0127] After feature extraction, there are usually several fully connected layers that integrate the previously extracted features for the final classification task.
[0128] Each neuron in a fully connected layer is connected to all neurons in the previous layer and is responsible for mapping features to categories.
[0129] Output layer:
[0130] Depending on the task requirements, the output layer can be a softmax function, used for multi-class classification problems, which transforms the output of the previous layer into the probability distribution of each class.
[0131] Backpropagation and optimization:
[0132] The loss function (such as cross-entropy loss) is calculated by comparing the predicted results with the true labels.
[0133] The process of updating network parameters using gradient descent or other optimization algorithms to reduce the loss function value is called backpropagation.
[0134] Training and validation:
[0135] The network learns iteratively on the training set while using a validation set to evaluate performance and prevent overfitting.
[0136] Through multiple iterations, the network gradually learns to recognize specific patterns and objects from images.
[0137] Through the above process, CNN can automatically learn and extract useful features from the original image, and perform image recognition and classification based on these features.
[0138] like Figure 2 As shown, the present invention provides an image recognition system based on deep learning, including an acquisition module, a calculation module, a determination module, a preprocessing module, and a recognition module;
[0139] The acquisition module is used to acquire the images that need to be recognized;
[0140] The calculation module is used to calculate the preprocessing coefficients of the image to be recognized;
[0141] The determination module is used to determine the preprocessing strategy for the image to be recognized based on the preprocessing coefficients;
[0142] The preprocessing module is used to preprocess the images to be recognized based on the preprocessing strategy to obtain preprocessed images;
[0143] The recognition module is used to input preprocessed images into a deep learning-based image recognition model for recognition.
[0144] The foregoing has shown and described the basic principles and main features of the present invention and its advantages. It will be apparent to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects. The scope of the present invention is defined by the appended claims rather than the foregoing description. Therefore, it is intended that all changes falling within the meaning and scope of the equivalents of the claims be included within the present invention, and no reference numerals in the claims should be regarded as limiting the scope of the claims.
[0145] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.
Claims
1. A deep learning-based image recognition method, characterized in that, include: S1, acquire the image to be recognized; S2, calculate the preprocessing coefficients of the image to be recognized; S3, determine the preprocessing strategy for the image to be recognized based on the preprocessing coefficients; S4, based on the preprocessing strategy, preprocess the image to be recognized to obtain the preprocessed image; S5, input the preprocessed image into a deep learning-based image recognition model for recognition; The calculation process for the preprocessing coefficients includes: The first step is to convert the image to grayscale to obtain a grayscale image. The second step is to obtain the computation region in the grayscale image; The third step is to calculate the preprocessing coefficients of the image to be identified based on the computational region; Obtain the computation region in the grayscale image, including: The images that need to be identified are divided into Two areas of equal size; Based on the distance from the center of the image to be identified, from near to far, calculate the region judgment value for each region sequentially. Starting from the third region judgment value, perform the following processing: Determine the value of the (k+1)th region. Does it meet the requirements? If so, then The corresponding region is used as the calculation region. and These represent the judgment values for the k-th and (k-1)-th regions, respectively; k is greater than or equal to 2. If not, check if k is equal to If not, continue calculating the region judgment value for the next region; if yes, take the region with the largest region judgment value as the calculation region. The formula for calculating the region judgment value is: ； This represents the judgment value of the k-th region obtained by calculating the k-th region. This represents the set of pixels in the k-th region. express The total number of pixels in the image. express The grayscale value of pixel i in the image; The preprocessing coefficients for the image to be recognized are calculated based on the computational region, including: Calculate the preprocessing coefficients using the following formula: ； Indicates the preprocessing coefficients. This represents the number of pixels in the calculation area whose grayscale value is greater than the average grayscale value of all pixels in the calculation area. This represents the number of pixels in image P obtained after processing the pixels in the computational region using the NLM algorithm, whose grayscale value is greater than the average grayscale value of all pixels in P. This represents the maximum grayscale value of the pixel in P. M1 represents the total number of pixels with grayscale value j in P, and M2 represents the total number of pixels in P. This represents the maximum grayscale value in the calculation area. M4 represents the number of pixels with a grayscale value of k in the calculation area. and These represent the first weight and the second weight, respectively.
2. The image recognition method based on deep learning according to claim 1, characterized in that, The image to be recognized is converted to grayscale to obtain a grayscale image, including: The image to be identified is converted to grayscale using a weighted average method to obtain a grayscale image.
3. The image recognition method based on deep learning according to claim 2, characterized in that, The image to be identified is converted to grayscale using a weighted average method to obtain a grayscale image, including: The following formula is used to calculate the grayscale image for each pixel in the image to be recognized: ； This represents the grayscale value of pixel b. , and These represent the red, green, and red component values of pixel b in the RGB color space, respectively. , and These are the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient, respectively.
4. The image recognition method based on deep learning according to claim 1, characterized in that, The preprocessing strategy for determining the image to be recognized is based on the preprocessing coefficients, including: If the preprocessing coefficient is greater than the preset coefficient threshold, the first preprocessing method is adopted as the preprocessing strategy for the image to be identified. If the preprocessing coefficients are less than or equal to the preset coefficient threshold, then the second preprocessing method is used as the preprocessing strategy for the image to be identified.
5. The image recognition method based on deep learning according to claim 4, characterized in that, The first preprocessing method includes: Let NP represent the image that needs to be preprocessed. For NP, the NLM algorithm and wavelet denoising algorithm are used to preprocess NP to obtain images NP1 and NP2 respectively. The formula for obtaining the preprocessed image PP is as follows: ； This represents the grayscale value of pixel b in image PP. and These represent the gray values of pixels in images NP1 and NP2 that have the same coordinates as pixel b in image PP. This indicates the set acquisition coefficient. .
6. A deep learning-based image recognition system, characterized in that, It includes an acquisition module, a calculation module, a determination module, a preprocessing module, and a recognition module; The acquisition module is used to acquire the images that need to be recognized; The calculation module is used to calculate the preprocessing coefficients of the image to be recognized; The determination module is used to determine the preprocessing strategy for the image to be recognized based on the preprocessing coefficients; The preprocessing module is used to preprocess the images to be recognized based on the preprocessing strategy to obtain preprocessed images; The recognition module is used to input the preprocessed image into a deep learning-based image recognition model for recognition. Obtain the computation region in the grayscale image, including: The images that need to be identified are divided into Two areas of equal size; Based on the distance from the center of the image to be identified, from near to far, calculate the region judgment value for each region sequentially. Starting from the third region judgment value, perform the following processing: Determine the value of the (k+1)th region. Does it meet the requirements? If so, then The corresponding region is used as the calculation region. and These represent the judgment values for the k-th and (k-1)-th regions, respectively; k is greater than or equal to 2. If not, check if k is equal to If not, continue calculating the region judgment value for the next region; if yes, take the region with the largest region judgment value as the calculation region. The formula for calculating the region judgment value is: ； This represents the judgment value of the k-th region obtained by calculating the k-th region. This represents the set of pixels in the k-th region. express The total number of pixels in the image. express The grayscale value of pixel i in the image; The preprocessing coefficients for the image to be recognized are calculated based on the computational region, including: Calculate the preprocessing coefficients using the following formula: ； Indicates the preprocessing coefficients. This represents the number of pixels in the calculation area whose grayscale value is greater than the average grayscale value of all pixels in the calculation area. This represents the number of pixels in image P obtained after processing the pixels in the computational region using the NLM algorithm, whose grayscale value is greater than the average grayscale value of all pixels in P. This represents the maximum grayscale value of the pixel in P. M1 represents the total number of pixels with grayscale value j in P, and M2 represents the total number of pixels in P. This represents the maximum grayscale value in the calculation area. M4 represents the number of pixels with a grayscale value of k in the calculation area. and These represent the first weight and the second weight, respectively.