A method and system for removing image bubble noise based on dual-channel features

Abstract

The present invention discloses a method and system for removing bubble noise from an image based on dual-channel features, and relates to the field of image processing technology. An improved neural network model is constructed, wherein the dual channels include a first channel for extracting color features related to image recognition, and a second channel for extracting bubble interference features; a feature processing unit is used to superimpose the bubble interference features and the color features related to image recognition and then extract them; a picture without interference and with concentration annotations and a picture partially with bubbles are obtained as interference-free images and bubble interference images, and the improved neural network model is trained and tested to obtain a trained improved neural network model; an image to be processed is obtained and input into the trained improved neural network model, and after processing such as convolution masking, the bubble noise is removed, and the image is expanded into a vector form after convolution, and the final output node is set to 1 to output the result of target recognition of the input image.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a method and system for removing image bubble noise based on dual-channel features. Background Art

[0002] Existing image recognition technologies often experience reduced recognition accuracy when processing images subject to specific interference. For example, in industrial inspection, medical image processing, and everyday visual recognition tasks, interference such as bubbles, reflections, and water stains in the image can affect the classification model's ability to correctly identify image targets.

[0003] Convolutional neural networks (CNNs) commonly used in image processing are mostly trained without noise interference, failing to fully consider the impact of interfering objects on model feature extraction, resulting in unstable model performance in interference environments. Even if images with interference are added for training, the model still cannot fully extract effective information from interference and non-interference. At the same time, no method for correctly processing interference information is subsequently added. Therefore, using this mixed training method has increased the accuracy of the model to a certain extent, but the accuracy is still very poor on unfamiliar datasets.

[0004] For example, during microalgae cultivation, bubbles significantly interfere with concentration monitoring. Bubble generation is closely related to processes such as aeration and stirring. These bubbles are randomly distributed within the culture medium and continuously generate and burst. The reflection and refraction effects of these bubbles create uneven brightness areas, disrupting color consistency in the image and increasing image noise, making color and brightness-based feature extraction more difficult. Furthermore, bubbles can cause local dynamic changes in the image, affecting the model's prediction accuracy.

[0005] In recent years, researchers have proposed some methods to alleviate bubble interference, reducing the interference of bubbles on image features through image filtering and denoising techniques. However, such algorithms are usually used for images with more obvious themes and richer color channels of feature information. If they are applied to target recognition in images with relatively uniform colors and unclear bubble colors, the accuracy and robustness of the model are poor. Summary of the Invention

[0006] In order to overcome the shortcomings of the above-mentioned prior art in the recognition of image targets with relatively uniform colors and unclear bubble colors, the main purpose of the present invention is to provide an image bubble noise removal method and system based on dual-channel features.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a method for removing bubble noise from an image based on dual-channel features, specifically comprising:

[0008] An improved CNN-based neural network model was constructed, comprising a first channel, a second channel, a plurality of sequentially connected feature processing units, a mask module, and a fully connected layer; the first channel was used to extract color features related to image recognition, and the second channel was used to extract bubble interference features; the feature processing unit comprised four convolutional layers and a feature superposition extraction module, which was used to superimpose the bubble interference features and color features related to image recognition before extracting them; the mask module was used to mask the output of the feature processing unit, and the fully connected layer was used to process the masked feature expansion and output the target recognition result;

[0009] Obtaining pictures without interference and with labeled information and some pictures with bubbles as non-interference images and bubble-interference images, constructing training sets, inputting them into the first channel and the second channel respectively, training the improved neural network model, and obtaining a trained improved neural network model;

[0010] Obtain the improved neural network model after training of the image to be processed and output the corresponding target recognition result.

[0011] Preferably, the convolution kernel of each convolution layer in each of the feature processing units is used for bubble interference feature extraction and color feature extraction related to image recognition.

[0012] Preferably, the step of obtaining the trained improved neural network model for inputting the image to be processed and outputting the corresponding target recognition result comprises the following steps:

[0013] Inputting the image to be processed into the improved neural network model to obtain color features and bubble interference features related to image recognition;

[0014] The color features related to image recognition and the extracted bubble interference features are input into a plurality of feature processing units connected in sequence for feature extraction, wherein:

[0015] In each of the feature processing units, four convolution layers are used to sequentially perform convolution processing on the input, and a filter is used to extract features from the convolved input to obtain re-extracted convolution input features. The re-extracted convolution input features are element-wise added to obtain superimposed features, and then features are extracted again to obtain intermediate enhanced bubble interference features.

[0016] generating a low-value matrix having the same size as the bubble interference feature according to the intermediate enhanced bubble interference feature;

[0017] Overlaying the generated low-value matrix onto the bubble interference feature, performing masking processing to obtain masked features, performing superimposed feature processing on the masked features and color features related to image recognition to remove bubble noise, and obtaining a bubble noise removal result;

[0018] The bubble noise removal result is expanded into a vector form using a fully connected layer, and the output node of the fully connected layer is set to 1 to output the corresponding target recognition result.

[0019] A system for removing image bubble noise based on dual-channel features, comprising:

[0020] A model improvement module, configured to construct an improved CNN-based neural network model, comprising a first channel, a second channel, a plurality of sequentially connected feature processing units, a mask module, and a fully connected layer; the first channel is configured to extract color features related to image recognition, and the second channel is configured to extract bubble interference features; the feature processing unit comprises four convolutional layers and a feature superposition extraction module, the feature superposition extraction module being configured to superimpose bubble interference features and color features related to image recognition before extracting them; the mask module being configured to mask the output of the feature processing unit, and the fully connected layer being configured to process the masked feature expansion and output the target recognition result;

[0021] A model training module is used to obtain images without interference and with labeled information and some images with bubbles as interference-free images and bubble-containing interference images, construct a training set, input them into the first channel and the second channel respectively, train the improved neural network model, and obtain a trained improved neural network model;

[0022] The bubble noise removal module obtains the image to be processed and inputs the trained improved neural network model, and outputs the corresponding target recognition result.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] 1. The dual-channel input of the present invention allows for simultaneous processing of interference-free images and images containing bubble interference, extracting pure color features and bubble interference features, respectively. By overlapping the feature maps of the dual-channel input data and superimposing the feature maps of the two channels, the neural network can better identify and distinguish bubble noise from the target to be identified. The bubble interference features are masked using a low-value matrix to reduce the weight of the interference area, thereby reducing the impact of bubble noise in subsequent processing. The method for improving the accuracy of online monitoring in an interference environment can effectively improve the fitting accuracy in unfamiliar and complex environments.

[0025] 2. This method employs feature superposition and re-extraction of post-convolution features, thereby addressing, to a certain extent, the requirement for high pixel consistency between the original image and the image with interference. Requiring only two images of the same concentration reduces the complexity of data acquisition and annotation. By dividing the convolution kernel into two parts, bubble interference features and color features can be more accurately extracted. Finally, a regression operation is performed on the output layer, allowing for fine-tuning of the prediction results to obtain a higher-quality image with bubble noise removed.

[0026] 3. The present invention targets concentration fitting models based on machine vision, improves fitting accuracy by processing images, and reduces the difficulty and cost of data acquisition.

[0027] 4. The present invention can effectively improve the sensitivity of partial convolution to areas that do not belong to the main features, ensuring the effective extraction of the main features and the accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The drawings described herein are used to provide further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute improper limitations on the present application.

[0029] Figure 1 is a heat map of the original image and the degree of attention paid by the visualization model to the original image in the embodiment of the present invention; wherein, Figure 1 a is the heat map of the original image, Figure 1 b is a heat map of the degree of attention paid by the basic model to the original image. Figure 1 c is a heat map of the degree of attention paid by the model to the original image after the introduction of the attention mechanism. Figure 1 d is a heat map of the degree of attention paid by the model of the present invention to the original image;

[0030] Figure 2 is a heat map of the attention level of the visualization model after the original image and the initial layer and four pooling layers in the model architecture in the embodiment of the present invention are processed; wherein, Figure 2 a is the heat map of the original image, Figure 2 b is the heat map after initial layer processing, Figure 2 c is the heat map after the first pooling layer processing, Figure 2 d is the heat map after processing by the second pooling layer, Figure 2 e is the heat map after processing by the third pooling layer, Figure 2 f is the heat map after processing by the fourth pooling layer;

[0031] Figure 3 It is a schematic diagram of the process structure of the present invention. DETAILED DESCRIPTION

[0032] During microalgae cultivation, bubbles significantly interfere with concentration monitoring. Bubble generation is closely related to processes such as aeration and stirring. These bubbles are randomly distributed within the culture medium and continuously generate and collapse. The reflection and refraction effects of these bubbles create uneven brightness areas, disrupting color consistency in the image and increasing image noise, making color- and brightness-based feature extraction more difficult. Furthermore, bubbles can cause local dynamic changes in the image, affecting the model's prediction accuracy.

[0033] In recent years, researchers have proposed some methods to alleviate bubble interference, reducing the interference of bubbles on image features through image filtering and denoising techniques. However, such algorithms are usually used for images with more obvious themes and richer feature information color channels. When applied to the concentration identification of microalgae solutions with more uniform colors and less obvious bubble colors, the robustness of the model still needs further optimization and evaluation.

[0034] This hybrid training approach improves model accuracy to a certain extent, but accuracy remains very poor on unfamiliar datasets. In particular, bubbles significantly interfere with concentration monitoring during microalgae cultivation. Bubble generation is closely related to processes such as aeration and stirring. These bubbles are randomly distributed within the culture medium and continuously generate and burst. The reflection and refraction effects of the bubbles create uneven brightness regions, disrupting image color consistency and increasing image noise, making color- and brightness-based feature extraction more difficult. Furthermore, bubbles can cause local dynamic changes in the image, thus affecting the model's prediction accuracy. In recent years, researchers have proposed methods to mitigate bubble interference, using image filtering and denoising techniques to reduce the impact of bubbles on image features. However, such algorithms are typically used for images with distinct themes and rich feature color channels. Applying them to microalgae solution concentration recognition, where the color is more uniform and the bubble color is less pronounced, requires further optimization and evaluation to ensure robustness.

[0035] The existing technology is specifically shown as follows:

[0036] (1) Microalgae identification and biomass prediction based on convolutional neural network algorithm: The article explains the feasibility of convolutional neural network in fitting the concentration of microalgae solution, but does not consider the interference of bubbles. The introduction of bubbles will cause great interference to the feature extraction of the model, resulting in a decrease in recognition accuracy.

[0037] (2) Microalgae concentration prediction method and system based on big data: This patent proposes a method to establish a prediction model based on historical sampling and environmental data. However, this method of using data to build a prediction model requires a large amount of data support, and for certain specific culture environments, parameter measurement requires the support of a large number of sensors. Prediction with the help of machine vision does not require a large amount of data and is easy to obtain.

[0038] (3) A bubble recognition image processing method for bubble flow in a gas-liquid reactor: This patent proposes a method for shadow monitoring and bubble compensation using background difference method. However, this method requires the background and the picture with bubbles to be differentiated, which is easily affected by the inconsistent matching of the original image pixels, resulting in a decrease in accuracy, and has certain requirements on the contrast between the background and bubbles.

[0039] (4) FFDNet: Toward a Fast and Flexible Solution for CNN based ImageDenoising: Most of the existing denoising methods are mainly aimed at noise removal in target detection and object recognition applications. The data source targets detected by these models are generally rich in color. However, in the field of microalgae concentration prediction, the color of the detected target is relatively uniform and single, and the introduced bubbles will produce different color states as the concentration of the algae solution changes. Therefore, the effect of introducing an additional denoising module to remove bubbles is not obvious. In order to overcome the above problems, the present invention proposes an image bubble noise removal method based on dual-channel features, see Figure 3 , specifically including:

[0040] An improved CNN-based neural network model was constructed, comprising a first channel, a second channel, a plurality of sequentially connected feature processing units, a mask module, and a fully connected layer; the first channel was used to extract color features related to image recognition, and the second channel was used to extract bubble interference features; the feature processing unit comprised four convolutional layers and a feature superposition extraction module, which was used to superimpose the bubble interference features and color features related to image recognition before extracting them; the mask module was used to mask the output of the feature processing unit, and the fully connected layer was used to process the masked feature expansion and output the target recognition result;

[0041] Obtaining pictures without interference and with labeled information and some pictures with bubbles as non-interference images and bubble-interference images, constructing training sets, inputting them into the first channel and the second channel respectively, training the improved neural network model, and obtaining a trained improved neural network model;

[0042] Obtain the improved neural network model after training of the image to be processed and output the corresponding target recognition result.

[0043] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0044] Example 1:

[0045] This example uses an image of microalgae as an example. To overcome the shortcomings of the existing technology, a method for removing bubble noise from a uniform color image based on overlapping dual-channel feature maps and performing masking is proposed. The specific details are as follows:

[0046] 1. Dataset Construction: Place the collected images of microalgae without interference and with concentration annotations, and some images of microalgae with bubbles, into two different folders. Annotate the folders.

[0047] 2. Model Input Channel Design: Modify the input layer structure of the neural network model, expanding the single-channel input to a dual-channel input. The two types of data mentioned above, interference-free images and images with bubble interference, are input into the model through different input channels to effectively distinguish the data sources.

[0048] 3. Feature extraction module optimization: After the initial convolutional layers for the model's interference-free and bubble-interference images, a feature map overlay and extraction module was added. This module extracts bubble interference features and color features related to microalgae concentration, allowing for the extraction of bubble and concentration features from the feature map.

[0049] As the model is exposed to real interference and background, it will be trained so that the convolution kernel in the model architecture has a certain degree of recognition of color features and bubble interference features.

[0050] 4. Introducing a masking operation: After the feature extraction module, a masking operation is added to mask the extracted bubble interference features, reducing their weight influence in the model prediction process. This effectively reduces the negative impact of bubble interference on concentration prediction accuracy.

[0051] 5. Model Structure Adjustment and Optimization: Within the model structure, a feature extraction and masking module was inserted every four convolutional layers as a feature processing unit to continuously suppress bubble interference. Furthermore, the number of nodes in the model's final output layer was set to 1, converting the model from a classification model to a regression model to achieve accurate fitting and continuous numerical prediction of microalgae concentration.

[0052] 6. Model training and generalization capability improvement: The improved model is trained to significantly improve its generalization capability and prediction accuracy for bubble interference conditions without reducing the concentration prediction accuracy under normal conditions.

[0053] Table 1 Performance of the basic model and two improved methods on two unfamiliar datasets

[0054]

[0055] Beneficial effects achieved: (1) Two models formed by adding the basic model and the attention mechanism mentioned in the existing literature and the method proposed in the present invention to the basic model were verified on two unfamiliar datasets. The verification results are shown in Table 1. The two unfamiliar datasets include: dataset a is pictures with bubbles at concentrations encountered during model training, and dataset b is pictures with bubbles at concentrations not encountered during model training. From the perspective of various evaluation parameters of the model, regardless of whether the model has learned the concentration gradients appearing in the unfamiliar dataset, feature map superposition and mask processing show higher prediction accuracy.

[0056] (2) The heat map after the model’s attention to the original image area is visualized as follows Figure 1 As shown, Figure 1 a is the heat map of the original image, Figure 1 b is a heat map of the degree of attention paid by the basic model to the original image. Figure 1 c is a heat map of the degree of attention paid by the model to the original image after the introduction of the attention mechanism. Figure 1 d is a heat map of the model's attention to the original image. By comparing the original image and the heat maps of each model, it can be seen that the masked model pays more attention to non-bubble areas, that is, areas containing more concentration information, which is why the model has higher accuracy.

[0057] (3) Draw a heat map of the attention level of the visual model after the initial layer and the four pooling layers in the model architecture, such as Figure 2 As shown, Figure 2 a is the heat map of the original image, Figure 2 b is the heat map after initial layer processing, Figure 2 c is the heat map after the first pooling layer processing, Figure 2 d is the heat map after processing by the second pooling layer, Figure 2 e is the heat map after processing by the third pooling layer, Figure 2 f is the heatmap after the fourth pooling layer. It can be seen that as the number of layers increases, the attention paid to the bubble area gradually decreases, and the mask module plays a role.

[0058] Example 2:

[0059] The microalgae concentration prediction method is based on masking to remove bubble noise in uniform color images based on overlapping dual-channel feature maps, which specifically includes the following steps

[0060] The first step is to construct the dataset, including a set of non-interference images and a set of bubble-interference images:

[0061] 1000 images of microalgae concentration without bubble interference were collected, and each image was accurately labeled with concentration, ranging from 0.1g / L to 1.0g / L, with an interval of 0.1g / L.

[0062] 500 images of microalgae concentrations containing different numbers and sizes of bubbles were collected, also with concentration annotations, covering the same concentration range.

[0063] Modify the input layer of the neural network model and expand the single-channel input to a dual-channel input; the interference-free image is input through channel A, and the bubble-interferenced image is input through channel B.

[0064] A feature map superposition and extraction module is added after the initial convolutional layer to extract bubble interference features and microalgae concentration-related color features from the feature map.

[0065] A masking operation is added after the feature extraction module to mask the extracted bubble interference features and reduce their weight to 10% of the original value.

[0066] In the model, a feature extraction and masking module was inserted every four convolutional layers as a feature processing unit. The final output layer node was set to 1, converting the model into a regression model. The model was trained using both a set of unobstructed images and a set of images with bubbles. During training, the ratio of unobstructed images to images with bubbles was 2:1. Cross-validation and early stopping were used to prevent overfitting. The training cycle was 100 epochs, with an initial learning rate of 0.001, which was then reduced by half every 20 epochs.

[0067] Concentration prediction accuracy under normal working conditions: On the interference-free image set, the average absolute error of the model-predicted concentration is 0.02 g / L, and the accuracy rate reaches 95%.

[0068] Generalization capability for working conditions with bubble interference: On a set of images with bubble interference, the model's average absolute error in predicted concentration was 0.04 g / L, and the accuracy was increased to 90%, an improvement of 15 percentage points compared to the model that did not use dual-channel feature map overlapping mask processing.

[0069] Overall model performance: Combining the two data sets, the model's overall prediction accuracy reached 92%, meeting the needs of real-time monitoring of microalgae concentrations.

[0070] Through the above examples, it can be seen that the proposed method effectively reduces the impact of bubble interference on the prediction of microalgae concentration, and improves the generalization ability and prediction accuracy of the model under complex working conditions.

[0071] Example 3:

[0072] A method for removing bubble noise in a uniform color image based on mask processing based on overlapping dual-channel feature maps specifically includes the following:

[0073] Collect and annotate microalgae concentration data, totaling 1,000 images.

[0074] The 520 images of undisturbed microalgae concentrations, with concentrations marked in the range of 100-1000 mg / L, were stored in the folder "without bubbles." The 480 images of microalgae concentrations with bubbles, with concentrations marked in the same range, were stored in the folder "with bubbles."

[0075] The folders are marked as "no bubbles" and "with bubbles", and each picture is named in the format of "concentration_XX mg / L_serial number.jpg" for subsequent tracking and analysis.

[0076] Based on the ResNet18 architecture, the input layer structure was modified from a single-channel input to a dual-channel input. The first channel inputs "no bubble" images, and the second channel inputs "bubble" images.

[0077] Add a feature map stacking and extraction module after the initial convolutional layer of the model.

[0078] A masking operation is added after the feature extraction module. The bubble interference weights calculated from the mask matrix will affect the prediction and are updated at each iteration. A feature extraction and masking module is inserted after every four convolutional layers as a feature processing unit. The number of nodes in the final output layer is set to 1, converting the model from a classification model to a regression model, outputting a continuous value of microalgae concentration.

[0079] The Adam optimizer was used with an initial learning rate of 0.001 and a batch size of 32. Training was performed for 50 epochs. Cross-validation was performed, with the data split into a training set (80%) and a test set (20%) to verify model performance. The model's predictive accuracy was evaluated on the test set containing bubbles, using the root mean square error (RMSE) as the metric, with the goal of keeping the RMSE below 10 mg / L.

[0080] After training is complete, the model is evaluated and found to be:

[0081] On the “no bubbles” image set, the model prediction error is 5 mg / L, RMSE.

[0082] On the "bubble" image set, the model prediction error is 8 mg / L, RMSE, showing strong robustness.

[0083] Further observation of the model's prediction effect when there are more bubbles shows that 80% of the predicted concentration errors are controlled within 10 mg / L, of which 10% of the sample concentrations exceed this range.

[0084] Through the above implementation plan, a model was constructed that can effectively deal with bubble interference and predict microalgae concentration, providing accurate technical support for microalgae cultivation, monitoring and related research.

[0085] It should be noted that, in the present invention, relational terms such as first and second, etc. are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include", "comprises", or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or apparatus that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article, or apparatus.

[0086] The above embodiments are merely examples of the present invention and do not limit the scope of protection of the present invention. Any designs that are identical or similar to the present invention fall within the scope of protection of the present invention.

Claims

1. A method for removing bubble noise from an image based on dual-channel features, characterized in that: The following steps are involved: An improved CNN-based neural network model was constructed, comprising a first channel, a second channel, a plurality of sequentially connected feature processing units, a mask module, and a fully connected layer; the first channel was used to extract color features related to image recognition, and the second channel was used to extract bubble interference features from the image; the feature processing unit comprised four convolutional layers and a feature superposition extraction module, which was used to superimpose the bubble interference features and color features related to image recognition before extracting them; the mask module was used to mask the output of the feature processing unit, and the fully connected layer was used to process the masked feature expansion and output the target recognition result; Obtaining pictures without interference and with labeled information and some pictures with bubbles as non-interference images and bubble-interference images, constructing training sets, inputting them into the first channel and the second channel respectively, training the improved neural network model, and obtaining a trained improved neural network model; Obtain the trained improved neural network model for the image to be processed and output the corresponding target recognition result; The step of obtaining the trained improved neural network model for inputting the image to be processed and outputting the corresponding target recognition result comprises the following steps: Inputting the image to be processed into the improved neural network model to obtain color features and bubble interference features related to image recognition; The color features related to image recognition and the extracted bubble interference features are input into a plurality of feature processing units connected in sequence for feature extraction, wherein: In each of the feature processing units, four convolution layers are used to sequentially perform convolution processing on the input, and a filter is used to extract features from the convolved input to obtain re-extracted convolution input features. The re-extracted convolution input features are element-wise added to obtain superimposed features, and then features are extracted again to obtain intermediate enhanced bubble interference features. generating a low-value matrix having the same size as the bubble interference feature according to the intermediate enhanced bubble interference feature; Overlaying the generated low-value matrix onto the bubble interference feature, performing masking processing to obtain masked features, performing superimposed feature processing on the masked features and color features related to image recognition to remove bubble noise, and obtaining a bubble noise removal result; The bubble noise removal result is expanded into a vector form using a fully connected layer, and the output node of the fully connected layer is set to 1 to output the corresponding target recognition result.

2. The image bubble noise removal method based on dual-channel features according to claim 1, characterized in that: The convolution kernel of each convolution layer in each of the feature processing units is used for bubble interference feature extraction and color feature extraction related to image recognition.

3. A system for removing image bubble noise based on dual-channel features, characterized in that: include: A model improvement module, configured to construct an improved CNN-based neural network model, comprising a first channel, a second channel, a plurality of sequentially connected feature processing units, a mask module, and a fully connected layer; the first channel is configured to extract color features related to image recognition, and the second channel is configured to extract bubble interference features; the feature processing unit comprises four convolutional layers and a feature superposition extraction module, the feature superposition extraction module being configured to superimpose bubble interference features and color features related to image recognition before extracting them; the mask module being configured to mask the output of the feature processing unit, and the fully connected layer being configured to process the masked feature expansion and output the target recognition result; A model training module is used to obtain images without interference and with labeled information and some images with bubbles as interference-free images and bubble-containing interference images, construct a training set, input them into the first channel and the second channel respectively, train the improved neural network model, and obtain a trained improved neural network model; The bubble noise removal module obtains the image to be processed and inputs the trained improved neural network model, and outputs the corresponding target recognition result; The step of obtaining the trained improved neural network model for inputting the image to be processed and outputting the corresponding target recognition result comprises the following steps: Inputting the image to be processed into the improved neural network model to obtain color features and bubble interference features related to image recognition; The color features related to image recognition and the extracted bubble interference features are input into a plurality of feature processing units connected in sequence for feature extraction, wherein: In each of the feature processing units, four convolution layers are used to sequentially perform convolution processing on the input, and a filter is used to extract features from the convolved input to obtain re-extracted convolution input features. The re-extracted convolution input features are element-wise added to obtain superimposed features, and then features are extracted again to obtain intermediate enhanced bubble interference features. generating a low-value matrix having the same size as the bubble interference feature according to the intermediate enhanced bubble interference feature; Overlaying the generated low-value matrix onto the bubble interference feature, performing masking processing to obtain masked features, performing superimposed feature processing on the masked features and color features related to image recognition to remove bubble noise, and obtaining a bubble noise removal result; The bubble noise removal result is expanded into a vector form using a fully connected layer, and the output node of the fully connected layer is set to 1 to output the corresponding target recognition result.

4. A computer-readable storage medium, characterized in that The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 2 is implemented.

5. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method according to any one of claims 1 to 2 is implemented.

Images