A convolution calculation method for improving the recognition accuracy of fuzzy biological images

Abstract

The present invention relates to a convolution calculation method for improving the recognition accuracy of fuzzy biological images, comprising: obtaining a fuzzy biological image; constructing an optimal convolution operation model, wherein the optimal convolution operation model comprises: obtaining all weight values ​​in a convolution kernel in a deep learning model, obtaining a weight average value through the weight value; determining a feature extraction factor according to the weight average value, introducing the feature extraction factor into the convolution operation model of the convolution kernel, and obtaining an optimal convolution operation model; inputting the fuzzy biological image into the optimal convolution operation model for feature extraction and convolution operation, and obtaining a feature map. The present invention is based on improving the convolution calculation of feature extraction in the original convolutional neural network (CNN), realizing pixel difference enhancement, thereby improving the recognition accuracy of the convolutional neural network model for fuzzy biological images.

Description

Technical Field

[0001] The present invention relates to the technical field of outdoor biological recognition, and in particular to a convolution calculation method for improving the recognition accuracy of fuzzy biological images. Background Art

[0002] Driven by deep learning, the technology of blurred target recognition has made significant progress. First, the EHDC framework extracts small target features through a hybrid dilated convolutional network, expands the perception field without increasing the algorithm calculation amount, and learns spatial semantic features through an adaptive correlation matrix to compensate for the lack of target features, thereby improving the recognition accuracy of small underwater blurred targets. Secondly, for the problem of target recognition in entangled light quantum imaging, a method based on a two-stream feature fusion network is proposed. Entangled photon pairs are prepared through spontaneous parametric down-conversion and quantum imaging is realized, and then a two-stream feature fusion convolutional network is designed to identify the reconstructed quantum image. In addition, for rotationally blurred biological images, a new rotation invariant target recognition method is proposed. By constructing a model containing three measurement layers, the correct target recognition rate of blurred biological images is improved.

[0003] In the field of space target recognition, the reference images are generated by three-dimensional models and simulated images, and the generative adversarial network GAN is used to improve the correctness, accuracy and efficiency of space target recognition based on ground observation. In addition, the fuzzy facial expression recognition system based on convolutional neural network (CNN) achieves high-accuracy facial expression recognition through CNN training with different depths. In the recognition of motion-blurred Chinese character coding targets CCTs, deep learning and transfer learning are used to reduce the number of required samples and training time, and improve the recognition accuracy. Finally, for the problem of infrared small target detection IRSTD, a new infrared shape network ISNet is used to extract the precise shape information of the target through specific edge blocks and bidirectional attention aggregation blocks, which has a significant effect on infrared target detection with low signal-to-noise ratio and low contrast.

[0004] In summary, deep learning technology is increasingly widely used in the field of blurred target recognition. Through different network structures and learning strategies, it can effectively improve recognition performance and provide strong technical support for future research and application. However, the above technologies are all optimizing feature extraction strategies and completing feature information, and do not solve the problem of small differences between blurred targets and backgrounds. It is urgent to propose a convolution calculation method for improving the accuracy of blurred biological image recognition, which can amplify the difference between blurred targets and backgrounds by modifying the convolution kernel weights. Summary of the invention

[0005] The convolution kernels of existing deep convolutional neural networks often find it difficult to effectively capture the features of the target when facing blurry biological images in the wild, which in turn affects the recognition accuracy, especially when performing target recognition under undesirable conditions such as blur or overexposure. In order to solve this problem, the present invention proposes a convolution calculation method for improving the recognition accuracy of blurry biological images, which improves the classification accuracy by implementing the function of pixel difference enhancement in a convolutional neural network (CNN).

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A convolution calculation method for improving the accuracy of fuzzy biological image recognition, comprising:

[0008] S1. Acquire a blurred biological image;

[0009] S2. Constructing an optimal convolution operation model, wherein constructing the optimal convolution operation model includes: obtaining all weight values ​​in the convolution kernel in the deep learning model, and obtaining a weight average value through the weight values; determining a feature extraction factor according to the weight average value, and introducing the feature extraction factor into the convolution operation model of the convolution kernel to obtain an optimal convolution operation model;

[0010] S3. Input the blurred biological image into the optimal convolution operation model to perform feature extraction and convolution operation to obtain a feature map.

[0011] Optionally, obtaining the weighted average value includes:

[0012] All the obtained weight values ​​are summed, and the summed weight value is divided by the weight quantity to obtain the weight average value.

[0013] Optionally, determining the feature extraction factor according to the weighted average value includes:

[0014] The weighted average value is compared with the weight value, and when the weighted average value is greater than the weight value, a first feature extraction factor is obtained, and when the weighted average value is less than the weight value, a second feature extraction factor is obtained.

[0015] Optionally, the first feature extraction factor is greater than the second feature extraction factor, and both the first feature extraction factor and the second feature extraction factor are greater than 1.

[0016] Optionally, obtaining the optimal convolution operation model includes:

[0017] The feature extraction factor is multiplied by the weight value inside the convolution kernel in the convolution operation model to obtain the optimal convolution operation model.

[0018] Optionally, the method for obtaining the convolution operation model of the convolution kernel is:

[0019] y=wx+b

[0020] Among them, y is the output value of the convolution kernel after the convolution operation, w is the weight value inside the convolution kernel, x is the input feature or data, and b is the bias term.

[0021] Optionally, the method for obtaining the optimal convolution operation model is:

[0022] y=w i *(p i )x i +b

[0023] Among them, y is the output value of the convolution kernel after the convolution operation, w is the weight value inside the i-th convolution kernel, and x i is the input feature or data, b is the bias term, p i is the feature extraction factor.

[0024] Optionally, the method further comprises: setting a time period, and when the time reaches the time period, repeating step S2.

[0025] The beneficial effects of the present invention are:

[0026] The present invention is based on improving the convolution calculation of feature extraction in the original convolutional neural network (CNN) to achieve pixel difference enhancement, thereby improving the recognition accuracy of the convolutional neural network model for fuzzy biological images.

[0027] The present invention improves the sensitivity of traditional convolution kernels to pixel differences and improves recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0029] Figure 1 A flow chart of a convolution calculation method for improving the recognition accuracy of fuzzy biological images according to an embodiment of the present invention;

[0030] Figure 2 Schematic diagram of pixel difference at the same position after being processed by a convolution kernel according to an embodiment of the present invention;

[0031] Figure 3 A schematic diagram of a training data set according to an embodiment of the present invention;

[0032] Figure 4 Schematic diagram of the accuracy of the improved network and their corresponding original versions on the validation set according to the embodiment of the present invention; wherein, Figure 4 (a) is a schematic diagram of the accuracy of the AlexNet network and their corresponding original versions on the validation set. Figure 4 (b) is a schematic diagram of the accuracy of the VGG-16 network and their corresponding original versions on the validation set. Figure 4 (c) is a schematic diagram of the accuracy of the ResNet-18 network and their corresponding original versions on the validation set;

[0033] Figure 5 Schematic diagram of the influence of different p values ​​on network classification accuracy in an embodiment of the present invention; wherein, Figure 5 (a) is a schematic diagram showing the effect of different p values ​​on the classification accuracy of the network in the AlexNet network. Figure 5 (b) is a schematic diagram showing the effect of different p values ​​on the classification accuracy of the network in the ResNet network. Figure 5 (c) Schematic diagram of the impact of different p values ​​on network classification accuracy in the VGGnet network;

[0034] Figure 6 The four evaluation indicators of the convolution operation of VGG-16 and VGG-16+ improved in the embodiment of the present invention are compared on the Oxford-flower dataset;

[0035] Figure 7 The four evaluation indicators of the convolution operation of ResNet-18 and ResNet-18+ improved in the embodiment of the present invention are compared on the Oxford-flower dataset;

[0036] Figure 8 The recognition accuracy of VGG-16+ improved convolution operation and ResNet-18+ improved convolution operation on Oxford-flower with different p values ​​in the embodiment of the present invention;

[0037] Fig. 9 The experimental results of setting different T values ​​using VGGNet-16+ improved convolution operations on Tensorflow-flower in an embodiment of the present invention are shown. DETAILED DESCRIPTION

[0038] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0040] like Figure 1 As shown, the present invention discloses a convolution calculation method for improving the accuracy of fuzzy biological image recognition, including: S1. Obtaining a fuzzy biological image; S2. Constructing an optimal convolution operation model, constructing the optimal convolution operation model includes: obtaining all weight values ​​in the convolution kernel in the deep learning model, and obtaining the weight average through the weight value; determining the feature extraction factor according to the weight average, introducing the feature extraction factor into the convolution operation model of the convolution kernel, and obtaining the optimal convolution operation model; S3. Inputting the fuzzy biological image into the optimal convolution operation model for feature extraction and convolution operation, and obtaining a feature map, specifically:

[0041] The present invention discloses a convolution calculation method for improving the accuracy of fuzzy biological image recognition. The operation improves the feature extraction capability by enhancing the convolution calculation method strategy of feature distance. The present invention improves the traditional convolution operation (Formula 1) and introduces a feature extraction factor (p). The factor is dynamically adjusted according to the convolution kernel weight, thereby generating a new operation (Formula 2).

[0042] like Figure 2 As shown, for all weight values ​​in the convolution kernel, the arithmetic mean is calculated; for each weight value Wi in the convolution kernel, the corresponding Pi value is determined based on its comparison with (average); using the determined Pi value and the convolution kernel weight value Wi, the modified convolution operation (Formula 3) is performed;

[0043] y=wx+b Formula 1

[0044] y=w*p*x+b Formula 2

[0045] y=w i *(p i )x i +b Formula 3

[0046] Among them, y is the output value of the convolution kernel after the convolution operation, w is the weight value inside the convolution kernel, x is the input feature or data, b is the bias term (or intercept), which allows the model to have a non-zero baseline output even when all input features are zero, P is the introduced hyperparameter, i is the different weight values ​​in the convolution kernel, and i in Xi represents the input features or data at different positions in the input feature map (i=1, 2, 3...).

[0047] For an input feature map x and a convolution kernel w, where the convolution kernel has a certain size, the value of the output Y of the convolution operation at each position (i, j) can be expressed by the following mathematical expression:

[0048]

[0049] Here, X(i+u, j+v) represents the value of position (i+u, j+v) on the input feature map, W(u, v) represents the weight of position (u, v) on the convolution kernel, b is the bias term, and Y(i, j) is the value of position (i, j) on the output feature map.

[0050] This expression provides a way to calculate the value of each position in the output feature map Y by applying the convolution kernel W to the local area of ​​the input feature map X and then adding the bias term b. This operation is repeated on the entire input feature map to obtain the complete output feature map Y.

[0051] In CNN, there is a significant mapping relationship between input features and convolution kernel weights. This mapping relationship is established through convolution operations and is the basis for CNN to perform effective feature extraction and learning. In the convolution layer, the convolution kernel slides (or convolves) on the input feature map, and the sum of the dot products of the convolution kernel and each local area of ​​the input feature map is calculated. This process produces a new feature map (also called an activation map or feature map), which contains information about the interaction between the input feature map and the convolution kernel. Therefore, it can be said that there is a direct mapping relationship between the convolution kernel weights and the input features. The mapping relationship between input features and convolution kernel weights is the core mechanism of feature extraction and learning of convolutional neural networks. In this way, CNN can automatically learn effective feature representations.

[0052] The following is the mathematical expression of the modified convolution operation:

[0053]

[0054]

[0055]

[0056] Operation steps:

[0057] 1. For all weight values ​​in the convolution kernel, calculate their arithmetic mean.

[0058] 2. For each weight value Wi in the convolution kernel, determine the corresponding Pi value based on its comparison with (average). There are two Pi values: p and 2-p.

[0059] 3. Using the determined Pi value and convolution kernel weight value Wi, perform the modified convolution operation (Formula 3).

[0060] Weighted average calculation formula:

[0061]

[0062] Specific application: In the field of outdoor biological identification, heavy rain, sandstorms, different lighting angles, and extreme temperatures may affect the clarity of images acquired by the light sensor of the shooting equipment. At this time, this research can solve the accuracy challenges of the recognition algorithm based on convolutional neural network technology.

[0063] It can solve problems in biometric applications in harsh and extreme environments, such as blurred images and videos due to weather and lighting, insufficient feature information collected, or poor image quality, which affect subsequent recognition and classification during fixed-point camera observations and tracking of rare wild animals.

[0064] Based on this method, the present invention transforms the standard AlexNet, VGG-16, and ResNet-18 models, and evaluates the recognition accuracy of the transformed models on multiple data sets through experiments. In addition, the experiment also explores the impact of different cycle parameters T and adjustment strategy P values ​​on model performance, in order to find a better model configuration.

[0065] Get all the weight values ​​in the convolution kernel in the deep learning model, including: Convolution kernel weight acquisition: In the process of training a deep learning model, you can programmatically grab (i.e. extract and view) the convolution kernel weights. In PyTorch, each layer of the model is a module, and its weights are directly accessed.

[0066] Obtaining the weight average includes: summing up all the weight values ​​obtained, dividing the summed weight value by the weight quantity, and obtaining the weight average, specifically:

[0067] The convolution kernel designed by the present invention adjusts the parameter (P) by comparing the difference between each weight weight in the convolution kernel and the weight average value, so that the convolution kernel can enhance or suppress the influence of specific pixels in the receptive field while retaining the image structure;

[0068] The weight average is calculated as follows: For all weight values ​​in the convolution kernel, calculate their arithmetic mean.

[0069] Determining the feature extraction factor according to the weight average value includes: comparing the weight average value with the weight value, when the weight average value is greater than the weight value, obtaining the first feature extraction factor, when the weight average value is less than the weight value, obtaining the second feature extraction factor, the first feature extraction factor is greater than the second feature extraction factor, and the first feature extraction factor and the second feature extraction factor are both greater than 1, specifically:

[0070] The adjustment of parameter P is based on the following logic: if the weight value of a convolution kernel is greater than the average weight of the convolution kernel, it indicates that the weight value of the convolution kernel may play a very important role in extracting the target features. In this case, by multiplying the weight value of the convolution kernel by a larger parameter value P1 (P1>1), both values ​​of parameter P are hyperparameters that can be optimized through experiments. Hyperparameters are parameters whose values ​​are set before starting the learning process, rather than parameter data obtained through training. Parameters (P) are adjusted by comparing the difference between each weight weight in the convolution kernel and the weight average.

[0071] The above two feature extraction factors should not be greater than 1. The convolution kernel weights that are less than the average weight are multiplied by a small parameter value p (p<1), and the convolution kernel weights that are greater than the average weight are multiplied by a larger parameter value 2-p to achieve the effect of differentiation, so as to increase the influence of these pixels in the subsequent feature map. On the contrary, if the value of the convolution kernel weight is less than the average weight of the convolution kernel, the weight values ​​of these convolution kernels are more likely to represent background or unimportant features, so their influence is reduced by multiplying them by a small parameter value P2 (P2<1). It is worth noting that both the values ​​of parameters P1 and P2 are hyperparameters that can be optimized through experiments.

[0072] like Figure 6-8 As shown, regarding the selection of P value:

[0073] Ablation experiments were designed. Under different P value settings, the improved convolution operation of VGG-16+ showed obvious changes in recognition accuracy on the Oxford-flower dataset. For example, when the p value is 1, it means that the model does not use the improved convolution operation model. As the P value increases, the recognition accuracy of the model gradually increases, reaches a peak, and then begins to gradually decrease. This shows that there is an optimal P value range that can achieve the best performance of the model. Similarly, the performance of the improved convolution operation of ResNet-18+ also shows the characteristics of being sensitive to the P value. Within a specific P value range, the performance of the model is optimal; beyond this range, the performance decreases.

[0074] In order to verify the effectiveness of the fuzzy target recognition method based on deep learning, the present invention selected multiple public data sets for testing. These data sets include but are not limited to flower classification data sets and self-constructed field data sets. These data sets cover different target types and background environments, and can comprehensively evaluate the recognition ability of the model.

[0075] like Figure 3As shown in the figure, the dataset contains 4 types of pictures: clear pictures, Gaussian noise pictures, highly exposed pictures, and grayscale pictures. Gaussian noise, exposure, and grayscale pictures are synthesized by code from the original images in the dataset to simulate the wild environment. The sample proportion of each type of picture is the same.

[0076] When using the modified convolution kernel for network model training, the present invention first optimizes the training process of the network model. In order to enable the network to better learn how to enhance pixel differences, the present invention sets a period parameter T, which is used to ensure that the parameters of the network convolution kernel are modified once every T cycles. The modification strategy is the method proposed by the present invention. The reason is that compared with directly modifying the randomly generated convolution kernel weights, modifying the trained convolution kernel weights, the network will be encouraged to discover and strengthen those weights that have an important influence on classification decisions, thereby improving the recognition accuracy of blurred biological images. In addition, the present invention also explores different initialization methods and adjustment strategies for T and P values ​​to evaluate their impact on model performance.

[0077] like Fig. 9 As shown, the selection of T in the study is:

[0078] In order to evaluate the impact of different T values ​​on the network, the present invention designs an ablation experiment, in which the p value and other variable parameters are fixed, and different T values ​​are set to use VGGNet-16+ modified convolution operation model for training and verification on the same data set. The results show that a reasonably selected T value can indeed improve the recognition accuracy of the model. The reason for this result is that when the modified convolution operation begins, the current weight of the convolution kernel has an important impact on the effect of the modified convolution operation. When designing the modified convolution operation model, the present invention assumes that there is a strong mapping relationship between the convolution kernel weight and the feature, but this mapping relationship requires multiple rounds of learning to achieve. At the same time, due to the deep depth of the VGGNet network, when T = 0 and T = 64, the network has not been fully learned, and the mapping relationship between the convolution kernel weight and the feature is not as strong as when T = 100 and T = 150. When the improved convolution operation model starts and the convolution kernel weight is reasonable, the main factor affecting the accuracy becomes the execution frequency of the improved convolution operation. If the frequency is too high, the network will not have time to refit, and the frequency will be too low. The improved convolution operation will have less effect on the network and will not achieve the desired effect. Therefore, the accuracy of T = 200 is lower than that of T = 150.

[0079] like Figure 4As shown in (a)-4(c), next, the present invention selects AlexNet, VGG, ResNet and other models for training. AlexNet, VGG, ResNet are classic deep learning models that have achieved excellent performance in multiple image recognition tasks. The present invention introduces improved convolution kernels in the convolutional layers of AlexNet, VGG, and ResNet, and retrains the entire network. During the training process, the present invention pays special attention to the recognition performance of blurred or overexposed images to ensure that the network can still maintain high accuracy under these challenging conditions.

[0080] Finally, the present invention has conducted extensive tests on the improved network. The experimental results show that compared with the original network, the method of the present invention improves the recognition rate by at least 1%. The experimental results of the present invention depend on the selection of T value and P value. The experiments conducted by the present invention cannot represent the optimal experimental results, but the results have been able to show that the technology of the present invention can indeed improve the recognition accuracy of blurred targets. Through this method, the present invention can directly extract clearer features from the original image without relying on traditional image enhancement and deblurring processing steps, thereby improving the recognition accuracy. Of course, the technology of the present invention can also be used in combination with image enhancement, attention mechanism and other technologies to achieve a higher recognition rate.

[0081] In order to comprehensively evaluate the performance of the model, the experiment uses recognition accuracy as the evaluation indicator, that is, the proportion of targets correctly recognized by the model; in addition, the generalization ability of the model is also considered, that is, the performance of the model on unseen data.

[0082] like Figure 5 As shown in (a)-5(c), the experiment first preprocesses the data set, including operations such as cropping, scaling and normalization of the image to meet the input requirements of the model. Then, three deep learning models are selected for training according to the experimental requirements. The present invention uses three deep learning models, AlexNet, VGG-16, and ResNet-18, for fuzzy target recognition and comparative experiments. In the model training stage, the present invention simultaneously trains and verifies the original network model and the network model modified by the present invention, and fixes all random parameters to ensure the fairness and repeatability of the experiment. Specifically, on AlexNet, the present invention only modifies the first two layers of convolution, on VGG-16, the present invention only modifies the first 4 layers of convolution, and on ResNet-18, the present invention only modifies the first 5 layers of convolution. Finally, the recognition accuracy and other indicators of the model are recorded and analyzed.

[0083] In order to ensure the reliability of the experimental results, the experiment was statistically analyzed based on multiple repeated experiments to reduce the impact of accidental errors on the experimental results. At the same time, the experiment also compared the performance of different models and algorithms to demonstrate the advantages of the proposed method.

[0084] This experiment ensures the validity and repeatability of the experimental results by fixing random parameters through strict experimental settings. The performance of the fuzzy object recognition method based on deep learning is comprehensively evaluated by using multiple data sets and evaluation indicators.

[0085] The present invention is based on improving the convolution calculation of feature extraction in the original convolutional neural network (CNN) to achieve pixel difference enhancement. The present invention proposes a convolution calculation method for improving the accuracy of fuzzy biological image recognition. The present invention comprehensively trains and tests the improved network and their corresponding original versions on a custom data set. This experiment includes 100 training and verification cycles to ensure the reliability and stability of the results. After the transformation of AlexNet, it was observed that the model performance was improved compared with the original network, and the classification accuracy increased by 1%. Similarly, after the improved method was implemented on the VGG-16 network, the accuracy of the model was 1.3% higher than the original network. As for ResNet, the network performance was also improved through the improvement of the present invention, and the accuracy increased by 0.3% compared with the original network. The experimental results on Oxford Flowers show that the recognition accuracy on ResNet and VGGNet increased by 1.7% and 1.8% respectively, and other indicators also increased. These results show that the method of the present invention can improve the accuracy of image recognition on different deep learning architectures, thereby proving the universality of the present invention and its potential value in image recognition tasks.

[0086] In the wild, many environmental factors such as heavy rainfall, sandstorms, different lighting angles and extreme temperatures can interfere with the clarity of images captured by light sensors, which leads to challenges in the accuracy of recognition algorithms based on convolutional neural network technology. Existing methods are still insufficient in the recognition of blurred targets. This is mainly because traditional algorithms rely too much on image preprocessing and feature extraction steps, and their recognition capabilities will be greatly weakened in situations where target feature information is scarce or image quality is low. To this end, the present invention proposes an innovative technical solution. Based on the original convolution calculation method for feature extraction, a convolution calculation method for improving the accuracy of blurred biological image recognition is proposed, thereby improving the recognition accuracy of the convolutional neural network model for blurred biological images.

[0087] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A convolution calculation method for improving the accuracy of fuzzy biological image recognition, characterized in that: include: S1. Acquire a blurred biological image; S2. Constructing an optimal convolution operation model, wherein constructing the optimal convolution operation model includes: obtaining all weight values ​​in the convolution kernel in the deep learning model, and obtaining a weight average value through the weight values; determining a feature extraction factor according to the weight average value, and introducing the feature extraction factor into the convolution operation model of the convolution kernel to obtain an optimal convolution operation model; Determining the feature extraction factor according to the weighted average value includes: Compare the weighted average value with the weight value, and when the weighted average value is greater than the weight value, obtain a first feature extraction factor, and when the weighted average value is less than the weight value, obtain a second feature extraction factor; The first feature extraction factor is greater than the second feature extraction factor, and both the first feature extraction factor and the second feature extraction factor are greater than 1; Obtaining the optimal convolution operation model includes: Multiplying the feature extraction factor by the weight value inside the convolution kernel in the convolution operation model to obtain the optimal convolution operation model; S3. Input the blurred biological image into the optimal convolution operation model to perform feature extraction and convolution operation to obtain a feature map.

2. The convolution calculation method for improving the accuracy of fuzzy biological image recognition according to claim 1, characterized in that: Obtaining the weighted average value includes: All the obtained weight values ​​are summed, and the summed weight value is divided by the weight quantity to obtain the weight average value.

3. The convolution calculation method for improving the accuracy of fuzzy biological image recognition according to claim 1, characterized in that: The method for obtaining the convolution operation model of the convolution kernel is: y=wx+b Among them, y is the output value of the convolution kernel after the convolution operation, w is the weight value inside the convolution kernel, x is the input feature or data, and b is the bias term.

4. The convolution calculation method for improving the accuracy of fuzzy biological image recognition according to claim 1, characterized in that: The method for obtaining the optimal convolution operation model is: y=w i *(p i )x i +b Among them, y is the output value of the convolution kernel after the convolution operation, w is the weight value inside the i-th convolution kernel, xi is the input feature or data, b is the bias term, and pi is the feature extraction factor.

5. The convolution calculation method for improving the accuracy of fuzzy biological image recognition according to claim 1, characterized in that: The method further includes: setting a time period, and when the time reaches the time period, repeating step S2.

Images