Microbial microscopic image target recognition method based on wavelet enhanced convolutional neural network

Abstract

The application discloses a microorganism microscopic image target recognition method based on a wavelet enhanced convolutional neural network, and belongs to the technical field of image processing.The application obtains a microscopic image through a microorganism microscopic scanning device, adopts directional frame labeling to construct an image data set, and divides the image data set into a training set, a verification set and a test set; then the wavelet enhanced convolutional neural network is trained, verified and tested by using the above data set; finally, the network after the test is used to process an image to be recognized, and a microorganism recognition result is obtained.The application combines the wavelet enhancement and the convolutional neural network technology, and effectively improves the precision of microorganism microscopic image target recognition.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a microorganism microscopic image target recognition method based on a wavelet enhanced convolutional neural network. BACKGROUND

[0002] Microorganism microscopic image recognition is one of the core technologies of biomedical detection, environmental monitoring and disease diagnosis. With the development of computer vision and deep learning, automatic detection methods based on image analysis gradually replace traditional manual observation, significantly improving detection efficiency and accuracy. However, the complexity of microorganism samples (such as bacterial adhesion accumulation, morphological diversity) and the limitations of imaging conditions (such as low resolution, low contrast, background noise) are still technical bottlenecks for AI automated recognition of bacterial bodies.

[0003] In microorganism microscopic image processing, the traditional method is to rely on manual feature extraction (such as histogram equalization, threshold segmentation), and to classify bacteria by morphology after bacterial staining. This method is time-consuming and depends on the experience of the operator.

[0004] Although the existing deep learning model YOLOv11 can achieve automatic classification, it is limited by the fact that the receptive field of traditional convolution is limited to local regions. There are still bottlenecks in fine-grained feature extraction under dense small target detection (such as the adhesion of bacterial bodies in microorganism microscopic images) and complex background interference. The core module C3k2 of YOLOv11 is based on the CSP (Cross Stage Partial) architecture, which processes the input feature map by segmentation and combines multiple Bottleneck layers to achieve feature reuse. Its advantage is to reduce parameter redundancy while maintaining feature expression ability, but it has limited ability to capture long-distance dependencies. The fixed size of the traditional convolution kernel (such as 3x3) cannot cover the global context information of the bacterial morphology in the microorganism microscopic image, resulting in insufficient adhesion bacterial boundary segmentation accuracy, and the problem of low microorganism bacterial recognition accuracy. SUMMARY

[0005] In view of the above problems in the prior art, the microorganism microscopic image target recognition method based on a wavelet enhanced convolutional neural network provided by the present application solves the problem of low microorganism bacterial recognition accuracy in the prior art.

[0006] In order to achieve the above-mentioned application purposes, the technical scheme adopted by the present application is as follows: a microorganism microscopic image target recognition method based on a wavelet enhanced convolutional neural network, comprising the following steps:

[0007] S1, scanning a sample by a microorganism microscopic scanning device to obtain a microorganism microscopic image;

[0008] S2, labeling the microorganism microscopic image using a directional frame, and constructing each labeled microorganism microscopic image into a microorganism microscopic image dataset;

[0009] S3. Divide the microbial microscopic image dataset into training set, validation set and test set;

[0010] S4, sequentially using the training set, validation set, and test set to train, validate, and test the wavelet enhanced convolutional neural network to obtain a tested wavelet enhanced convolutional neural network;

[0011] S5. Use the tested wavelet enhanced convolutional neural network to process the microbial microscopic image to obtain the microbial identification result.

[0012] Furthermore, the wavelet enhanced convolutional neural network in S4 is specifically as follows: the C3K2 module in the original Yolov11 neural network is replaced with the wavelet enhanced convolution module WEConv to obtain a wavelet enhanced convolutional neural network.

[0013] Furthermore, the wavelet enhanced convolution module WEConv processes the input feature map in the following way:

[0014] A1. Use a wavelet filter bank to perform wavelet decomposition on the input feature map to obtain low-frequency components and high-frequency components;

[0015] A2, use 3×3 convolution layer to extract low-frequency features from low-frequency components;

[0016] A3, use 5×5 convolutional layer to extract high-frequency features from high-frequency components;

[0017] A4. Perform inverse wavelet transform on the high-frequency features and low-frequency features to obtain reconstructed features;

[0018] A5. Perform basic convolution on the input feature map to obtain basic features;

[0019] A6. Fuse the basic features with the reconstructed features to obtain a fused feature map;

[0020] A7. Determine whether the size of the fused feature map is equal to 2×2. If not, jump to step A1 and use the low-frequency component as the new input feature map in A1. If yes, output the fused feature map.

[0021] Furthermore, the wavelet filter bank in A1 includes: a low-pass filter, a horizontal high-pass filter, a vertical high-pass filter, and a diagonal high-pass filter;

[0022] The low-pass filter is: ;

[0023] The horizontal high-pass filter is: ;

[0024] The vertical high-pass filter is: ;

[0025] The diagonal high-pass filter is: ;

[0026] The high-frequency component in A1 includes a horizontal high-frequency component, a vertical high-frequency component and a diagonal high-frequency component.

[0027] Further, A1 includes the following sub-steps:

[0028] A11, convolve the input feature map with a low-pass filter to obtain a low-frequency component;

[0029] A12, convolve the input feature map with a horizontal high-pass filter to obtain a horizontal high-frequency component;

[0030] A13, convolve the input feature map with a vertical high-pass filter to obtain a vertical high-frequency component;

[0031] A14, convolve the input feature map with a diagonal high-pass filter to obtain a diagonal high-frequency component.

[0032] Further, the expression for obtaining the low-frequency component in A11 is:

[0033]

[0034] wherein X LL is the low-frequency component, f LL is the low-pass filter, is the convolution operation, and X is the input feature map.

[0035] The expression for obtaining the horizontal high-frequency component in A12 is:

[0036]

[0037] wherein X LH is the horizontal high-frequency component, f LH is the horizontal high-pass filter.

[0038] The expression for obtaining the vertical high-frequency component in A13 is:

[0039]

[0040] wherein X HL is the vertical high-frequency component, f HL is the vertical high-pass filter.

[0041] The expression for obtaining the diagonal high-frequency component in A14 is:

[0042]

[0043] wherein X HHFor diagonal high frequency components, f HH is a diagonal high-pass filter.

[0044] Further, A4 comprises the following sub-steps:

[0045] A41, convolve the low frequency feature with a low-pass transpose kernel to obtain a low frequency restored feature;

[0046] A42, convolve the horizontal high frequency feature with a horizontal transpose kernel to obtain a horizontal restored feature;

[0047] A43, convolve the vertical high frequency feature with a vertical transpose kernel to obtain a vertical restored feature;

[0048] A44, convolve the diagonal high frequency feature with a diagonal transpose kernel to obtain a diagonal restored feature;

[0049] A45, splice the low frequency restored feature, the horizontal restored feature, the vertical restored feature and the diagonal restored feature to obtain a reconstructed feature.

[0050] Further, the low-pass transpose kernel in A41 is: ;

[0051] The horizontal transpose kernel is: ;

[0052] The vertical transpose kernel is: ;

[0053] The diagonal transpose kernel is: .

[0054] Further, the formula for obtaining the low frequency restored feature in A41 is:

[0055]

[0056] wherein Z LL is the low frequency restored feature, is the low-pass transpose kernel, Y LL is the low frequency feature, is a convolution operation;

[0057] The formula for obtaining the horizontal restored feature in A42 is:

[0058]

[0059] wherein Z LH is the horizontal restored feature, is the horizontal transpose kernel, Y LH is the horizontal high frequency feature;

[0060] The formula for obtaining the vertical restored feature in A43 is:

[0061]

[0062] wherein, Z HL is a vertical reduction feature, is a vertical transpose kernel, Y HL is a vertical high frequency feature.

[0063] The formula of the diagonal reduction feature in A44 is:

[0064]

[0065] wherein, Z HH is a diagonal reduction feature, is a diagonal transpose kernel, Y HH is a diagonal high frequency feature.

[0066] Further, the convolution kernel size of the basic convolution in A5 is 5*5.

[0067] The beneficial effects of the present application are: the present application is aimed at the complexity of microbial samples (bacterial adhesion, morphological diversity) and imaging limitations (low resolution, noise, etc.), and through the wavelet enhanced convolutional neural network, the wavelet multi-scale decomposition characteristics can effectively extract different frequency features, break the limitations of traditional methods relying on artificial features, accurately capture bacterial details and contours, and adapt to complex imaging scenarios.

[0068] The present application solves the problem of "local receptive field, long distance dependence capture weak" of the Yolov11 model, and the wavelet decomposition can obtain global context information, combined with the convolution network, solves the defect that the traditional fixed size convolution cannot cover the global morphology of the bacteria, improves the adhesion bacterial boundary segmentation precision, strengthens the fine-grained feature extraction, and improves the microbial bacterial recognition precision. BRIEF DESCRIPTION OF DRAWINGS

[0069] Figure 1 It is a flow chart of a wavelet enhanced convolutional neural network microbial microscopic image target recognition method;

[0070] Figure 2 It is a structure diagram of a wavelet enhanced convolutional neural network;

[0071] Figure 3 It is a flow chart of a wavelet enhanced convolution module WEConv;

[0072] Figure 4 It is a recognition effect diagram of the Yolo11 original model;

[0073] Figure 5 It is a recognition effect diagram of the technical scheme of the present application. DETAILED DESCRIPTION

[0074] The specific embodiments of the present application are described below to facilitate the understanding of the present application for those skilled in the art, but it should be clear that the present application is not limited to the scope of the specific embodiments, and for those skilled in the art, it is obvious that various changes are within the spirit and scope of the present application defined and determined by the appended claims, and all the inventions utilizing the concept of the present application are within the scope of protection.

[0075] As shown in Figure 1 A microorganism microscopic image target recognition method of a wavelet enhanced convolutional neural network, comprising the following steps:

[0076] S1, scanning a sample by a microorganism microscopic scanning device to obtain a microorganism microscopic image;

[0077] S2, labeling the microorganism microscopic image by using a directional frame, and constructing each labeled microorganism microscopic image into a microorganism microscopic image dataset;

[0078] S3, dividing the microorganism microscopic image dataset into a training set, a verification set and a test set;

[0079] S4, training, verifying and testing the wavelet enhanced convolutional neural network by using the training set, the verification set and the test set in sequence to obtain a wavelet enhanced convolutional neural network after testing;

[0080] S5, processing a microorganism microscopic image to be recognized by using the wavelet enhanced convolutional neural network after testing to obtain a microorganism recognition result.

[0081] In the embodiment, the wavelet enhanced convolutional neural network in S4 is specifically: replacing a C3K2 module in an original Yolov11 neural network with a wavelet enhanced convolution module WEConv to obtain a wavelet enhanced convolutional neural network, as shown in Figure 2 .

[0082] In order to improve the ability of a deep learning model in microorganism microscopic image target recognition, the present application proposes a method of changing a Yolov11 model structure, replacing a C3k2 module with a wavelet enhanced convolution module WEConv, so as to realize the fusion of image multi-scale features and improve the precision of separating low-frequency background (such as slide staining impurities) and high-frequency details (such as bacterial edge) in a microorganism microscopic image.

[0083] In the embodiment, S1 is specifically: scanning a sample by a microorganism microscopic scanning device to obtain a high-definition microorganism microscopic image with an image size of 2976 pixels x 2976 pixels.

[0084] The professional personnel with microbiological testing experience will label the high-definition microbiological microscopic images obtained using an oriented bounding box (Oriented Bounding Box). The oriented bounding box is a bounding box annotation type, and the bounding box can be rotated to be more closely aligned with the detected object, and can better adapt to objects that are tilted or densely arranged.

[0085] The difference between the oriented bounding box and the ordinary bounding box is shown in Table 1.

[0086] Table 1

[0087]

[0088] The conventional CNN is limited by the size of the convolution kernel, and it is difficult to effectively capture global context information. Although some progress can be made by increasing the size of the convolution kernel, this usually leads to an explosive increase in the number of parameters and saturation of model performance. The wavelet-enhanced convolution module WEConv utilizes wavelet transform to expand the convolution receptive field through multi-frequency response and perform small-kernel convolution operations in different frequency ranges. Through wavelet decomposition, the model can capture low-frequency information in a larger range while avoiding over-parameterization of the model.

[0089] In this embodiment, the processing process of the wavelet-enhanced convolution module WEConv on the input feature map includes:

[0090] A1, wavelet decomposition of the input feature map is performed using a wavelet filter bank to obtain low-frequency components and high-frequency components;

[0091] A2, a 3x3 convolution layer is used to extract low-frequency features from the low-frequency components;

[0092] A3, a 5x5 convolution layer is used to extract high-frequency features from the high-frequency components;

[0093] A4, inverse wavelet transform is performed on the high-frequency features and the low-frequency features to obtain reconstructed features;

[0094] A5, basic convolution is performed on the input feature map to obtain basic features;

[0095] A6, the basic features and the reconstructed features are fused to obtain a fused feature map;

[0096] A7, it is judged whether the size of the fused feature map is equal to 2x2, if not, jump to step A1, and take the low-frequency components as the new input feature map in A1, if yes, output the fused feature map, as shown in Figure 3 .

[0097] In this embodiment, the wavelet filter bank in A1 includes a low-pass filter, a horizontal high-pass filter, a vertical high-pass filter, and a diagonal high-pass filter;

[0098] The low-pass filter is: , which is used to capture low-frequency information of the image (such as overall brightness and smooth areas), and the resolution is reduced to 1 / 4 of the original image;

[0099] The horizontal high-pass filter is: , which is used to extract high-frequency details in the horizontal direction (such as the left-right difference of vertical edges);

[0100] The vertical high-pass filter is: , which is used to extract high-frequency details in the vertical direction (such as the up-down difference of horizontal edges);

[0101] The diagonal high-pass filter is: , which is used to capture high-frequency changes in the diagonal direction (such as diagonal edges or texture details);

[0102] The high-frequency components in A1 include: horizontal high-frequency components, vertical high-frequency components, and diagonal high-frequency components.

[0103] The present application uses a predefined wavelet filter set ( ) to convolve the input, extract low-frequency components ( ) and high-frequency components ( ), and output the size of the input is half, while retaining the main outline covering the global morphology of the fungus respectively capturing horizontal, vertical and diagonal sensitive details, retaining wrinkle, crack and other detailed features.

[0104] In this embodiment, A1 includes the following steps:

[0105] A11, convolve the input feature map with a low-pass filter to obtain a low-frequency component;

[0106] A12, convolve the input feature map with a horizontal high-pass filter to obtain a horizontal high-frequency component;

[0107] A13, convolve the input feature map with a vertical high-pass filter to obtain a vertical high-frequency component;

[0108] A14, convolve the input feature map with a diagonal high-pass filter to obtain a diagonal high-frequency component.

[0109] In this embodiment, the expression of the low-frequency component obtained in A11 is:

[0110]

[0111] where X LL is the low-frequency component, f LL is the low-pass filter, is the convolution operation, and X is the input feature map;

[0112] The expression of the horizontal high frequency component in A12 is:

[0113]

[0114] wherein X LH is the horizontal high frequency component, f LH is a horizontal high pass filter;

[0115] The expression of the vertical high frequency component in A13 is:

[0116]

[0117] wherein X HL is the vertical high frequency component, f HL is a vertical high pass filter;

[0118] The expression of the diagonal high frequency component in A14 is:

[0119]

[0120] wherein X HH is the diagonal high frequency component, f HH is a diagonal high pass filter.

[0121] In the embodiment, A4 comprises the following steps:

[0122] A41, convolve the low frequency feature with a low pass transpose kernel to obtain a low frequency restored feature;

[0123] A42, convolve the horizontal high frequency feature with a horizontal transpose kernel to obtain a horizontal restored feature;

[0124] A43, convolve the vertical high frequency feature with a vertical transpose kernel to obtain a vertical restored feature;

[0125] A44, convolve the diagonal high frequency feature with a diagonal transpose kernel to obtain a diagonal restored feature;

[0126] A45, splice the low frequency restored feature, the horizontal restored feature, the vertical restored feature and the diagonal restored feature to obtain a reconstructed feature.

[0127] The application restores spatial resolution by transpose convolution, and ensures lossless reconstruction of information.

[0128] In the embodiment, the low pass transpose kernel in A41 is: ;

[0129] The horizontal transpose kernel is: ;

[0130] The vertical transpose kernel is: ;

[0131] The diagonal transpose kernel is: .

[0132] In this embodiment, the formula for obtaining the low-frequency restoration feature in A41 is:

[0133]

[0134] wherein Z LL is the low-frequency restoration feature, is the low-pass transpose kernel, Y LL is the low-frequency feature, is a convolution operation;

[0135] The formula for obtaining the horizontal restoration feature in A42 is:

[0136]

[0137] wherein Z LH is the horizontal restoration feature, is the horizontal transpose kernel, Y LH is the horizontal high-frequency feature;

[0138] The formula for obtaining the vertical restoration feature in A43 is:

[0139]

[0140] wherein Z HL is the vertical restoration feature, is the vertical transpose kernel, Y HL is the vertical high-frequency feature;

[0141] The formula for obtaining the diagonal restoration feature in A44 is:

[0142]

[0143] wherein Z HH is the diagonal restoration feature, is the diagonal transpose kernel, Y HH is the diagonal high-frequency feature.

[0144] In this embodiment, the convolution kernel size of the basic convolution in A5 is 5x5.

[0145] The present application adds the basic feature and the reconstruction feature element by element in A6 to obtain the fusion feature map.

[0146] As Figure 3As shown, the basic convolution is performed on the current feature map in the recursive process, the spatial local feature extraction is directly performed on the input feature map, the channel level dynamic scaling is performed, the important features are strengthened, the noise is suppressed, and the problem of limited feature expression ability caused by parameter sharing of the depth separable convolution is solved. The last basic convolution path (high frequency details) and the wavelet path (low frequency global) of the recursion are fused by element addition, that is, the spatial local detail features (edges and textures) of the basic convolution are retained, and the global features (such as contours) of the low frequency are captured, and the accuracy of identifying the microbial target is improved. The size of the wavelet convolution after recursion is continuously reduced, such as the size of 2x2 cannot be reduced any more, at this time, the receptive field of the small convolution kernel is expanded to the feature map of the complete size.

[0147] In the present embodiment, the labeled data set is divided into a training set, a validation set, and a test set in a ratio of 6:2:2. The training set and the validation set are input into the improved network model for iterative training to obtain the optimal model weight.

[0148] The present application successfully obtains the optimal model weight file by training the Yolo11 original model for 348 cycles. Subsequently, the weight file is applied to the target recognition task of the microbial microscopic image (image size is 2976x2976). Through model inference analysis, 626 gram-stained positive cocci are accurately identified in the image, and the recognition effect is as shown in Figure 4 .

[0149] The present application uses a wavelet-enhanced convolutional neural network to obtain the best model weight file through 192 cycles of training. The best model weight file is applied to the target recognition task of the same microbial microscopic image (image size is 2976x2976), and a total of 841 gram-stained positive cocci are identified, and the recognition result is as shown in Figure 5 .

[0150] The wavelet-enhanced convolutional neural network has higher convergence efficiency, and the improvement rate is 44.83%; at the same time, the recognition rate is also improved by 34.34%, as shown in Table 2.

[0151] Table 2

[0152]

[0153] The present application is aimed at the complexity of microbial samples (bacterial adhesion, morphological diversity) and imaging limitations (low resolution, noise, etc.), and through the wavelet-enhanced convolutional neural network, the wavelet multi-scale decomposition characteristics are used to effectively extract different frequency features, break the limitations of traditional methods relying on artificial features, accurately capture bacterial details and contours, and adapt to complex imaging scenarios.

[0154] The application makes up the problem that the Yolov11 and the like model is "local receptive field, long distance dependence captures weak", wavelet decomposition can obtain global context information, in combination with a convolution network, solves the defect that a traditional fixed size convolution cannot cover the global morphology of the bacteria body, improves the adhesion bacteria body boundary segmentation precision, strengthens the fine-grained feature extraction, and improves the microorganism bacteria body recognition precision.

[0155] The above is only the preferred embodiment of the present application and is not used to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement and the like within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for target recognition of microbial microscopic images using wavelet-enhanced convolutional neural networks, characterized in that: The following steps are involved: S1. Scan the sample using a microbial microscopy scanning device to obtain a microbial microscopic image; S2. Use oriented frames to annotate microbial microscopic images, and construct each annotated microbial microscopic image into a microbial microscopic image dataset; S3. Divide the microbial microscopic image dataset into training set, validation set and test set; S4, sequentially using the training set, validation set, and test set to train, validate, and test the wavelet enhanced convolutional neural network to obtain a tested wavelet enhanced convolutional neural network; S5, using the tested wavelet enhanced convolutional neural network to process the microbial microscopic image to be identified to obtain the microbial identification result; The wavelet enhanced convolutional neural network in S4 is specifically as follows: the C3K2 module in the original Yolov11 neural network is replaced with the wavelet enhanced convolution module WEConv to obtain a wavelet enhanced convolutional neural network; The processing process of the wavelet enhanced convolution module WEConv on the input feature map includes: A1. Use a wavelet filter bank to perform wavelet decomposition on the input feature map to obtain low-frequency components and high-frequency components; A2, use 3×3 convolution layer to extract low-frequency features from low-frequency components; A3, use 5×5 convolutional layer to extract high-frequency features from high-frequency components; A4. Perform inverse wavelet transform on the high-frequency features and low-frequency features to obtain reconstructed features; A5. Perform basic convolution on the input feature map to obtain basic features; A6. Fuse the basic features with the reconstructed features to obtain a fused feature map; A7. Determine whether the size of the fused feature map is equal to 2×2. If not, jump to step A1 and use the low-frequency component as the new input feature map in A1. If yes, output the fused feature map.

2. The microbial microscopic image target recognition method based on wavelet-enhanced convolutional neural network according to claim 1 is characterized in that: The wavelet filter group in A1 includes: a low-pass filter, a horizontal high-pass filter, a vertical high-pass filter and a diagonal high-pass filter; The low-pass filter is: ; The horizontal high-pass filter is: ; The vertical high-pass filter is: ; The diagonal high-pass filter is: ; The A1 high-frequency components include: horizontal high-frequency components, vertical high-frequency components and diagonal high-frequency components.

3. The microbial microscopic image target recognition method using wavelet-enhanced convolutional neural network according to claim 2 is characterized in that: A1 includes the following steps: A11. Use a low-pass filter to perform convolution operation on the input feature map to obtain the low-frequency component; A12. Use a horizontal high-pass filter to perform a convolution operation on the input feature map to obtain a horizontal high-frequency component; A13. Use a vertical high-pass filter to perform a convolution operation on the input feature map to obtain a vertical high-frequency component; A14. Use a diagonal high-pass filter to perform a convolution operation on the input feature map to obtain a diagonal high-frequency component.

4. The microbial microscopic image target recognition method using wavelet-enhanced convolutional neural network according to claim 3 is characterized in that: The expression for the low-frequency component obtained in A11 is: Among them, X LL is the low frequency component, f LL is a low-pass filter, is the convolution operation, X is the input feature map; The expression for the horizontal high-frequency component obtained in A12 is: Among them, X LH is the horizontal high frequency component, f LH is a horizontal high-pass filter; The expression of the vertical high-frequency component obtained in A13 is: Among them, X HL is the vertical high frequency component, f HL is a vertical high-pass filter; The expression of the diagonal high-frequency component obtained in A14 is: Among them, X HH is the diagonal high frequency component, f HH is a diagonal high-pass filter.

5. The microbial microscopic image target recognition method using wavelet-enhanced convolutional neural network according to claim 2, characterized in that: The A4 comprises the following steps: A41. Use a low-pass transposed kernel to perform convolution operation on the low-frequency features to obtain low-frequency restored features; A42, use the horizontal transpose kernel to perform convolution operation on the horizontal high-frequency features to obtain the horizontal restoration features; A43, use the vertical transpose kernel to perform convolution operation on the vertical high-frequency features to obtain the vertical restoration features; A44. Use the diagonal transpose kernel to perform convolution operation on the diagonal high-frequency features to obtain the diagonal restored features; A45. Concatenate the low-frequency restoration features, horizontal restoration features, vertical restoration features, and diagonal restoration features to obtain reconstructed features.

6. The microbial microscopic image target recognition method using wavelet-enhanced convolutional neural network according to claim 5 is characterized in that: The low-pass transpose core in the A41 is: ; The horizontal transposed kernel is: ; The vertical transposed kernel is: ; The diagonal transposed kernel is: .

7. The method for microbial microscopic image target recognition using a wavelet-enhanced convolutional neural network according to claim 5, wherein: The formula for obtaining the low-frequency restoration feature in A41 is: Among them, Z LL For low-frequency restoration characteristics, is the low-pass transposed kernel, Y LL is a low-frequency feature, is the convolution operation; The formula for obtaining the horizontal reduction feature in A42 is: Among them, Z LH is the horizontal reduction feature, is the horizontal transposed kernel, Y LH is the horizontal high-frequency feature; The formula for obtaining the vertical reduction feature in A43 is: Among them, Z HL is the vertical reduction feature, is the vertical transposed kernel, Y HL It is a vertical high-frequency feature; The formula for obtaining the diagonal restoration feature in A44 is: Among them, Z HH is the diagonal restoration feature, is the diagonal transposed kernel, Y HH is the diagonal high-frequency feature.

8. The microbial microscopic image target recognition method using wavelet-enhanced convolutional neural network according to claim 1 is characterized in that: The convolution kernel size of the basic convolution in A5 is 5×5.

Images