Intelligent ear disease identification method and system suitable for various types of equipment

Abstract

The invention discloses an ear disease intelligent identification method and system suitable for various devices, and the method comprises the steps: obtaining an ear lesion image data set, and carrying out the preprocessing of the ear lesion image data set, and obtaining a preprocessed image; constructing an intelligent identification model of ear diseases, wherein the intelligent identification model comprises a down-sampling module, a feature extraction module and an ear injury feature fusion module; performing down-sampling on the preprocessed image based on a down-sampling module to obtain a first feature map and a second feature map; based on a feature extraction module, performing high-level semantic feature extraction on the first feature map to obtain a first feature information map; performing low-level semantic feature extraction on the second feature map to obtain a second feature information map; and based on an ear injury feature fusion module, performing feature fusion on the first feature information graph and the second feature information graph to obtain a fused image. The fidelity rate of the fused image can be improved, global and local space information is captured at the same time, and the problem of low precision is relieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of ear disease recognition, and in particular relates to an ear disease intelligent recognition method and system applicable to multiple types of equipment. Background Art

[0002] The ear is one of the most important organs in the human body. It not only controls hearing, but also plays an important role in maintaining body balance. At the same time, the ear is also an organ that is often affected by diseases. If not detected and treated in time, ear diseases may lead to hearing loss and lifelong complications. Like other diseases, timely detection and appropriate treatment of ear diseases can help prevent worsening of the condition and avoid unnecessary side effects of certain drugs (such as broad-spectrum antibiotics). Compared with other diseases, ear disease examination is a simple and non-invasive process. In clinical practice, otoscopy is a commonly used ear examination method. However, due to differences in expertise and experience levels, misdiagnosis by medical professionals often occurs, especially in non-specialized clinics (pediatrics, general practitioners, emergency departments, primary care physicians, and telemedicine) and in areas with limited medical resources.

[0003] In order to alleviate the above challenges, it is urgent to develop a new generation of ear disease recognition devices that can provide timely, convenient and accurate recognition. Fortunately, the ear endoscope is commonly used for ear examination and is an easily replaceable device that can be easily operated without professional training. This provides the feasibility of integrating self-recognition capabilities into the device using emerging deep learning technologies. Although some studies have used deep learning models to intelligently recognize various types of ear endoscope images and achieved promising recognition results, these works often lack extensive data validation and only focus on model development and accuracy without considering the applicability of the model in actual scenarios. In fact, in addition to focusing on recognition performance, the intelligent recognition model of ear diseases should also balance the trade-offs between recognition accuracy, model parameter size, and model response speed. Clinical recognition often requires repeated observation of the affected ear canal area. Therefore, if a model can achieve extremely fast inference speed, it can allow individual users and otolaryngologists to conduct a thorough and in-depth scan of ear lesions in a real-time video format. This is more rigorous and scientific than relying solely on intelligent recognition based on a single static image, and is in line with the actual clinical recognition process. In addition, such models can be deployed on edge devices with limited computing power, such as low-end computers, smartphones, and tablets. Small model parameter size means minimal memory and storage requirements, enabling easy deployment on resource-constrained devices such as smartphones and embedded systems. Specifically, the key to using artificial intelligence to identify ear diseases lies in developing a system that has the above characteristics and is compatible with different devices. Such a system does not require expensive computing equipment or network services and can be reliably applied to a variety of local devices and scenarios, facilitating early detection and treatment of ear diseases for a wide range of patients. This in turn effectively reduces the overall impact and losses caused by ear diseases. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes an intelligent ear disease recognition method and system applicable to multiple types of devices to solve the problems existing in the above-mentioned prior art.

[0005] To achieve the above object, the present invention provides an intelligent ear disease recognition method applicable to multiple types of devices, comprising:

[0006] Acquire an ear lesion image dataset, and preprocess the ear lesion image dataset to obtain a preprocessed image;

[0007] Constructing an intelligent recognition model for ear diseases, wherein the intelligent recognition model comprises: a downsampling module, a feature extraction module and an ear injury feature fusion module;

[0008] Based on the downsampling module, downsampling the preprocessed image to obtain a first feature map and a second feature map;

[0009] Based on the feature extraction module, high-level semantic features are extracted from the first feature map to obtain a first feature information map; low-level semantic features are extracted from the second feature map to obtain a second feature information map;

[0010] Based on the ear injury feature fusion module, the first feature information map and the second feature information map are feature fused to obtain a fused image; based on the fused image, intelligent identification of ear diseases is achieved.

[0011] Preferably, the process of obtaining the first characteristic map and the second characteristic map includes:

[0012] Downsampling the preprocessed image multiple times and generating a downsampled image;

[0013] The downsampled image is downsampled in multiple stages to obtain a first feature map and a second feature map.

[0014] Preferably, the process of performing multi-stage downsampling on the downsampled image comprises:

[0015] Performing a second stage of downsampling on the downsampled image to obtain a second feature map;

[0016] Performing a third-stage downsampling on the second feature map to obtain a third-stage image;

[0017] The third stage image is downsampled in the fourth stage to obtain the first feature map.

[0018] Preferably, the process of obtaining the first characteristic information graph includes:

[0019] The first feature map is processed by a point-by-point convolution layer, a batch normalization layer, and a linear rectification function layer in sequence to obtain a first feature information map.

[0020] Preferably, the process of obtaining the second characteristic information graph includes:

[0021] The second feature map is processed by multiple average pooling, point-by-point convolution, regularization, efficient channel attention, batch normalization, and linear rectification function in sequence to obtain the second feature information map.

[0022] Preferably, the process of efficient channel attention processing includes:

[0023] The image input to the efficient channel attention is the efficient channel first image. After the efficient channel first image is processed by global average pooling and the first activation function layer in sequence, the university channel second image is obtained. The efficient channel first image and the university channel second image enter the multiplication layer to obtain the multiplication layer output image, which is the efficient channel attention output image.

[0024] Preferably, the process of obtaining the fused image includes:

[0025] Adding elements of the first feature information map and the second feature information map to obtain a first fused feature map;

[0026] Extracting a first weight and a second weight of the first fused feature map;

[0027] Based on the first multiplication layer, the first weight and the second feature information map are element-wise multiplied to calculate a first weight fusion feature map;

[0028] Based on the second multiplication layer, the second weight and the first feature information map are element-wise multiplied to calculate a second weight fusion feature map;

[0029] Based on the addition layer, the first weighted fusion feature map and the second weighted fusion feature map are element-wise added to obtain a fused image.

[0030] Preferably, the ear disease intelligent identification method further comprises:

[0031] Performing loss calculation on the second feature map to obtain a first loss value;

[0032] Performing loss calculation on the third-stage image to obtain a second loss value;

[0033] Performing loss calculation on the fused image to obtain a third loss value;

[0034] Based on the loss function, the first loss value, the second loss value, and the third loss value are calculated to obtain a final loss value for optimizing the intelligent recognition model.

[0035] In order to achieve the above technical objectives, the present invention further provides an ear disease intelligent recognition system applicable to multiple types of devices, which is used to implement any of the ear disease intelligent recognition methods applicable to multiple types of devices, and the system comprises:

[0036] An image acquisition module is used to acquire an ear lesion image dataset, and preprocess the ear lesion image dataset to obtain a preprocessed image;

[0037] A model building module, used to build an intelligent recognition model for ear diseases, wherein the intelligent recognition model includes: a downsampling module, a feature extraction module and an ear injury feature fusion module;

[0038] A downsampling module, used for downsampling the preprocessed image to obtain a first feature map and a second feature map;

[0039] A feature extraction module, configured to extract high-level semantic features from the first feature map to obtain a first feature information map; and extract low-level semantic features from the second feature map to obtain a second feature information map;

[0040] The ear injury feature fusion module is used to fuse the first feature information map and the second feature information map to obtain a fused image; based on the fused image, intelligent identification of ear diseases can be achieved.

[0041] The present invention also provides an electronic device, comprising: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, the method for intelligently identifying ear diseases applicable to multiple types of devices is implemented.

[0042] Compared with the prior art, the present invention has the following advantages and technical effects:

[0043] The present invention provides an intelligent ear disease recognition method applicable to multiple types of equipment, comprising: firstly acquiring an ear lesion image data set, and preprocessing the ear lesion image data set to obtain a preprocessed image; secondly constructing an intelligent ear disease recognition model, wherein the intelligent recognition model comprises: a downsampling module, a feature extraction module and an ear damage feature fusion module; further based on the downsampling module, downsampling the preprocessed image to obtain a first feature map and a second feature map; then based on the feature extraction module, performing high-level semantic feature extraction on the first feature map to obtain a first feature information map; performing low-level semantic feature extraction on the second feature map to obtain a second feature information map; finally, based on the ear damage feature fusion module, performing feature fusion on the first feature information map and the second feature information map to obtain a fused image; based on the fused image, intelligent ear disease recognition is achieved.

[0044] The present invention improves the fidelity of the fused image by using multi-level pixel dimensions and multi-weighted fusion factors before fusion of the image. The present invention can simultaneously capture global and local spatial information through the ear injury feature fusion module, and guide the network to focus on key areas in the feature map at each level, thereby alleviating the low precision problem. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0046] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0047] Figure 2 is a system schematic diagram of an embodiment of the present invention;

[0048] Figure 3 It is a unit-level framework diagram of an embodiment of the present invention;

[0049] Figure 4 It is a sub-unit level framework diagram of an embodiment of the present invention;

[0050] Figure 5 It is a framework diagram of the class header function model of an embodiment of the present invention;

[0051] Figure 6 A framework diagram of an efficient channel attention layer according to an embodiment of the present invention;

[0052] Figure 7 It is a framework diagram of the local-global spatial feature fusion unit of an embodiment of the present invention. DETAILED DESCRIPTION

[0053] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0054] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0055] Embodiment 1

[0056] like Figure 1 and Figure 3 As shown, this embodiment provides an intelligent ear disease recognition method applicable to multiple types of devices, including:

[0057] S1. Input an input image to be detected; wherein the input image is a pre-established ear lesion image dataset;

[0058] S2, extracting multi-class features from the input image to obtain a multi-class feature image, and then fusing the image after weight assignment to obtain a fused image; wherein the multi-class feature extraction includes: high-level semantic feature extraction and low-level semantic feature extraction, and the multi-class feature image includes: a first feature information graph after high-level semantic feature extraction and a second feature information graph after low-level semantic feature extraction;

[0059] Further, such as Figure 4 As shown, the S2 method includes:

[0060] S2.1, obtain an input image, normalize the input image that meets the preset conditions and generate it as a preprocessed image input S2.3, otherwise enter S2.2; wherein the preset condition is that the input image size is 3×224×224 dimensions, and the normalization is to use the average value [0.485, 0.456, 0.406] and the standard deviation [0.229, 0.224, 0.225] to normalize the three channels of the scaled input image;

[0061] S2.2, preprocessing the input image to obtain a preprocessed image;

[0062] Furthermore, the method of S2.2 includes:

[0063] S2.2.1. Scale each input image to 3×224×224 dimensions.

[0064] S2.2.2. The three channels of the scaled input image are normalized using the mean [0.485, 0.456, 0.406] and standard deviation [0.229, 0.224, 0.225] and generated as the preprocessed image.

[0065] S2.3, performing multiple downsampling on the preprocessed image and generating a downsampled image;

[0066] Furthermore, the S2.3 method includes:

[0067] S2.3.1, perform two 3×3 convolutions on the preprocessed image to obtain a 24×112×112-dimensional first downsampled image after downsampling twice;

[0068] S2.3.1 performs batch normalization on the first downsampled image and then performs rectification through a linear rectification function, and finally performs two maximum pooling operations to obtain a 24×56×56-dimensional downsampled image after two downsampling operations. The number of multiple downsampling operations in S2.3 is preferably 4 times, and the 4 downsampling operations are respectively 2 times of 3×3 convolution and 2 times of maximum pooling.

[0069] S2.4, after performing multi-stage downsampling on the downsampled image, the first feature map is finally obtained; wherein, the number of multi-stage downsampling in S2.4 is 3 times, the number of downsampling in each stage is 2, and a feature extraction is performed after each stage of downsampling.

[0070] Furthermore, the method of S2.4 includes:

[0071] S2.4.1, performing a second stage of downsampling on the downsampled image to obtain a second feature map; specifically, performing two downsamplings on the 24×56×56-dimensional downsampled image to obtain a second feature map of 48×28×28 dimensions;

[0072] S2.4.2, downsampling the second feature map in the third stage to obtain a third stage image; specifically, downsampling the second stage image of 48×28×28 dimensions twice to obtain a third stage image of 96×14×14 dimensions;

[0073] S2.4.3, downsample the third stage image to obtain a first feature map. Specifically, downsample the third stage image of 96×14×14 dimensions twice to obtain a first feature map of 192×7×7 dimensions.

[0074] S2.5, performing feature extraction on the first feature map to obtain a first feature information map after high-level semantic feature extraction;

[0075] Furthermore, the S2.5 method includes:

[0076] The first feature map is processed by the point-by-point convolution layer, batch normalization layer, and linear rectification function layer in sequence to obtain the first feature information map. Among them, the point-by-point convolution layer converts the 192×7×7-dimensional second feature map into 384×7×7 dimensions. Specifically, at the end of the best ear network model trunk, the first feature map output after the fourth stage of downsampling uses a convolution layer with a kernel size of 3x3 for the final high-level semantic feature extraction, and then batch normalization and ReLU activation function processing are performed to output the first feature information map.

[0077] S2.6, obtaining a second feature map obtained by downsampling in the second stage in S2.4, and performing feature extraction on the second feature map to obtain a second feature information map after low-level semantic feature extraction;

[0078] Furthermore, the S2.6 method includes: the second feature map obtained by the second stage of downsampling in S2.4 is the second feature map obtained in S2.4.1, and the second feature map is processed by multiple average pooling (Avg Pool), point-by-point convolution (PWConv), regularization (Drop Block), efficient channel attention (ECA), batch normalization (Batch Normal), and linear rectification function (ReLU) in sequence to obtain the second feature information map. The number of multiple average pooling in S2.6 is 4 times, and the second feature map of 48×28×28 dimensions is converted to 384×7×7 dimensions after 4 average pooling and one point-by-point convolution. Specifically, for the second feature map, it is not directly passed to the local global spatial feature fusion unit (LGSFF), but is processed through a separate branch path. First, the average pooling layer and the point-by-point convolution layer are applied to adjust the second feature map from 48×28×48 to 384×7×7. Then, the regularization layer is used to perform regularization on the feature map. After that, efficient channel attention (ECA) is applied to enable the network to pay more attention to the low-level feature information in the second feature map. Batch normalization and linear rectification function activation functions are applied to the second feature map features.

[0079] Furthermore, if Figure 6 As shown, the image input to the efficient channel attention is the first image of the efficient channel. The first image of the efficient channel is globally average pooled and then passes through the first activation function layer to obtain the second image of the university channel. The first image of the efficient channel and the second image of the university channel enter the multiplication layer, and the image output by the multiplication layer is the image output by the efficient channel attention.

[0080] S2.7. Perform feature fusion on the first feature information map and the second feature information map to obtain a fused image.

[0081] Furthermore, if Figure 7 As shown, the S2.7 feature fusion method includes:

[0082] S2.7.1. Add the elements of the first feature information map and the second feature information map to obtain a first fused feature map; specifically, add the elements of the first feature information map of 384×7×7 dimensions and the elements of the second feature information map of 384×7×7 dimensions to obtain a first fused feature map of 384×7×7 dimensions;

[0083] S2.7.2, extracting a first weight and a second weight of the first fused feature map;

[0084] Furthermore, the S2.7.2 method includes:

[0085] S2.7.2.1, subject the first fused feature map of 384×7×7 dimensions to grouped point-by-point convolution, batch normalization, and linear rectification function processing to obtain a second fused feature map of 384×7×7 dimensions;

[0086] S2.7.2.2. Concatenate the second fused feature map after the maximum pooling of 1×7×7 dimensions and the second fused feature map after the average pooling of 1×7×7 dimensions according to the channel dimension (cat) to obtain a third fused feature map of 2×7×7 dimensions;

[0087] S2.7.2.3, perform 3×3 convolution on the third fused feature map of 2×7×7 dimensions and input it into the multiplication layer with the second fused feature map of 384×7×7 dimensions to obtain a fourth fused feature map of 1×7×7 dimensions;

[0088] S2.7.2.4. Perform weight calculation on the fourth fused feature map of 1×7×7 dimensions to obtain a first weight and a second weight.

[0089] S2.7.3, performing element-wise multiplication of the first weight and the second feature information map based on the multiplication layer to obtain a first weight fusion feature map;

[0090] S2.7.4. Based on the multiplication layer, the second weight and the regularized first feature information map are element-wise multiplied to obtain a second weighted fusion feature map;

[0091] S2.7.5. Based on the addition layer, the first weighted fusion feature map and the second weighted fusion feature map are element-wise added to obtain a fused image.

[0092] In S2.7, the size of the first feature information map and the second feature information map are both 384×7×7; the local-global spatial feature fusion unit first performs an addition layer to add the elements of the first feature information map and the second feature information map, and then sequentially passes through grouped point-by-point convolution (GPWconv), batch normalization (BatchNormal), and linear rectification function (Relu) to generate the second fused feature map. In order to enhance the model's perception of spatial information in the input image and improve the model performance, the spatial attention module (Woo et al., 2018) is used to process the second fused feature map and obtain the third fused feature map. Then, the activation function (Sigmod) is calculated based on the third fused feature map to obtain the activation weights (first weight and second weight). Then, the first weight is multiplied by the second feature information map, and the second weight is multiplied by the first feature information map. Before multiplying the first feature information map by the second weight, the regularization (Dropblock) technique is applied to the first feature information map. Dropblock randomly discards some blocks, that is, groups of adjacent feature map units, instead of randomly discarding single neurons like Dropout, by forcing the model to use other non-discarded areas to learn image features; Dropblock provides stronger regularization, improving the robustness and generalization ability of the model. Finally, the first feature information map and the second feature information map multiplied by the corresponding weights are added to obtain the output, denoted as Out. The output dimension of EFLL, that is, the output dimension of the fused image, is the same as that of the first feature information map and the second feature information map, which is 384×7×7. The entire process of the local-global spatial feature fusion unit is summarized as formula (1), formula (2), and formula (3):

[0093]

[0094] SAout=Conv 3×3 (Cat(Maxpool(Fmerge);Averagepool(Fmerge)))(2)

[0095]

[0096] Where: Fmerge represents the second fusion feature map, Rule represents the linear rectification function, BN represents batch normalization, GPWconv represents grouped point-by-point convolution, Flow represents the second feature information map, represents the addition layer, Fhigh represents the first feature information map, SAout represents the third fusion feature map, Conv 3×3represents 3×3 convolution, Cat represents the concatenation of feature maps according to the channel dimension, Maxpool represents maximum pooling, Averagepool represents average pooling, Out represents the output of the local-global spatial feature fusion unit, Sigmod represents the activation function, represents a multiplication layer and DropBlock represents regularization.

[0097] Furthermore, the method also includes: S2.8, calculating the final loss value of the optimal ear net model based on the image and loss function obtained in any one or more steps of S2.1-S2.7.

[0098] Furthermore, the method of S2.8 includes:

[0099] S2.8.1. Calculate the loss of the second feature map to obtain a first loss value;

[0100] S2.8.2, calculate the loss of the third stage image to obtain a second loss value;

[0101] S2.8.3, calculating the loss of the fused image to obtain a third loss value;

[0102] S2.8.4. Based on the loss function, a final loss value calculation is performed on the first loss value, the second loss value, and the third loss value to obtain a final loss value for optimizing the optimal ear net model.

[0103] Among them, the first type of header function, the second type of header function, and the third type of header function all adopt a head-like function model, and the head-like function model adopts an input layer, a linear layer, a normalized reference function layer, and an output layer set in sequence, such as Figure 5 As shown. Specifically, multiple auxiliary classification heads are added to the optimal ear network model, and the first-class head function, the second-class head function, and the third-class head function are all auxiliary classification heads. Compared with a single loss, the use of multiple auxiliary classification heads introduces additional losses from the shallow layers of the network, which increases the gradient signal propagated during back propagation; this provides additional regularization, improves the robustness of the model, and helps the network better converge and optimize its parameters.

[0104] Furthermore, the method also includes S2.9, making predictions based on the results obtained by the prediction function through the third type header function.

[0105] S3. Output the fused image.

[0106] Beneficial effects of this embodiment:

[0107] This embodiment provides an intelligent ear disease recognition method applicable to multiple types of devices, constructs a large-scale data set, improves the richness of samples, performs multi-level class feature extraction on the input image, obtains multi-level class feature images, and then fuses them after weight assignment to obtain a fused image. The fidelity of the fused image is improved by using the multi-level pixel dimensions and multi-weight fusion factors before fusing the image.

[0108] This embodiment provides a novel local-global spatial feature fusion module that can capture global and local spatial information at the same time and guide the network to focus on key areas in feature maps at all levels, thereby alleviating the low precision problem. It is improved using local-global spatial feature fusion units and multiple auxiliary loss accumulation strategies to strike a balance between recognition performance, average FPS, and model parameter size.

[0109] In this embodiment, multiple class head function sub-units are added in the hidden layer as auxiliary classification heads to calculate the final fused image and the loss value of each downsampling stage. Compared with calculating a single loss, the use of multiple auxiliary classification heads will introduce additional losses from the shallow layers of the network, which will increase the gradient signal propagated during the back propagation process; this provides additional regularization, improves the robustness of the model, and helps the network better converge and optimize its parameters.

[0110] In this embodiment, the network uses multiple auxiliary classification heads for effective parameter optimization. With 0.77M parameters, the average frame rate of the best ear network model on the CPU reaches 80 frames per second. The system can be deployed on common cloud servers, mobile phones, tablets, and PC electronic devices. By operating a small electronic otoscope, users can use real-time video to fully scan and identify the ear canal, providing a new paradigm for ear endoscopes and other medical endoscopic image recognition applications.

[0111] This embodiment is applied to edge computing-based mobile terminals, cloud server-based mobile terminals, edge computing-based tablet terminals, and edge computing-based PC terminals. This embodiment designs a model that can be easily used in various real-world scenarios and develops a corresponding intelligent recognition model for ear lesions. To achieve this, the largest ear lesion image dataset ever was first established. Then, the lightweight network framework was used as the basic model, and a new image feature fusion module called ear lesion feature fusion module (ELFF) was proposed. This network achieves a good trade-off between recognition performance, inference speed, and model parameter size. This embodiment demonstrates ultra-fast inference speed and extremely small model parameter size, making it easy to deploy in different scenarios.

[0112] Importantly, this embodiment has excellent recognition performance for multiple types of ear diseases as well as normal ears, and has been validated by different populations of different genders, age groups, and clinical settings. Finally, an application was developed consisting of four versions: a mobile version based on edge computing, a mobile version based on a cloud server, a tablet version based on edge computing, and a PC version based on edge computing. The mobile version is applied to the mobile terminal, the tablet version is applied to the tablet, and the PC version is applied to the PC terminal, mainly helping ordinary users to perform self-identification and regular examinations. With the help of a compact electronic otoscope, users can identify their ears in time when they feel uncomfortable. At the same time, the Pad and PC versions can help healthcare professionals identify patients in various environments, including community screening, community clinics, resource-limited areas, and specialized ENT practices. This helps reduce misdiagnosis, improve recognition efficiency, enhance patient experience, and reduce the large amount of manpower and material costs associated with traditional recognition. In addition, a gradient CAM method is used to visualize the decision process of the optimal present system. This allows users and doctors to better understand the decision process of deep learning models and helps developers to more effectively verify, debug, and improve the behavior of models.

[0113] Embodiment 2

[0114] like Figure 2 As shown, this embodiment provides an ear disease intelligent recognition system applicable to multiple types of devices, which is used to implement the ear disease intelligent recognition method applicable to multiple types of devices described in any one of the first embodiments, and the system includes:

[0115] An image acquisition module is used to acquire an ear lesion image dataset, and preprocess the ear lesion image dataset to obtain a preprocessed image;

[0116] A model building module, used to build an intelligent recognition model for ear diseases, wherein the intelligent recognition model includes: a downsampling module, a feature extraction module and an ear injury feature fusion module;

[0117] A downsampling module, used for downsampling the preprocessed image to obtain a first feature map and a second feature map;

[0118] A feature extraction module, configured to extract high-level semantic features from the first feature map to obtain a first feature information map; and extract low-level semantic features from the second feature map to obtain a second feature information map;

[0119] The ear injury feature fusion module is used to fuse the first feature information map and the second feature information map to obtain a fused image; based on the fused image, intelligent identification of ear diseases can be achieved.

[0120] This embodiment has all the advantages of the first embodiment.

[0121] Implementation Three

[0122] This embodiment provides an electronic device, which includes: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, the intelligent ear disease recognition method applicable to multiple types of devices described in Embodiment 1 is implemented.

[0123] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. An intelligent ear disease recognition method applicable to multiple types of devices, It is characterized in that The following steps are involved: Acquire an ear lesion image dataset, and preprocess the ear lesion image dataset to obtain a preprocessed image; Constructing an intelligent recognition model for ear diseases, wherein the intelligent recognition model comprises: a downsampling module, a feature extraction module and an ear injury feature fusion module; Based on the downsampling module, downsampling the preprocessed image to obtain a first feature map and a second feature map; Based on the feature extraction module, high-level semantic features are extracted from the first feature map to obtain a first feature information map; low-level semantic features are extracted from the second feature map to obtain a second feature information map; Based on the ear injury feature fusion module, the first feature information map and the second feature information map are feature fused to obtain a fused image; based on the fused image, intelligent identification of ear diseases is achieved.

2. The method for intelligently identifying ear diseases applicable to multiple types of devices according to claim 1, It is characterized in that The process of obtaining the first feature map and the second feature map includes: Downsampling the preprocessed image multiple times and generating a downsampled image; The downsampled image is downsampled in multiple stages to obtain a first feature map and a second feature map.

3. The method for intelligently identifying ear diseases applicable to multiple types of devices according to claim 2, It is characterized in that The process of performing multi-stage downsampling on the downsampled image includes: Performing a second stage of downsampling on the downsampled image to obtain a second feature map; Performing a third-stage downsampling on the second feature map to obtain a third-stage image; The third stage image is downsampled in the fourth stage to obtain the first feature map.

4. The method for intelligently identifying ear diseases applicable to multiple types of devices according to claim 1, It is characterized in that The process of obtaining the first characteristic information graph includes: The first feature map is processed by a point-by-point convolution layer, a batch normalization layer, and a linear rectification function layer in sequence to obtain a first feature information map.

5. The method for intelligently identifying ear diseases applicable to multiple types of devices according to claim 1, It is characterized in that The process of obtaining the second characteristic information graph includes: The second feature map is processed by multiple average pooling, point-by-point convolution, regularization, efficient channel attention, batch normalization, and linear rectification function in sequence to obtain the second feature information map.

6. The method for intelligently identifying ear diseases applicable to multiple types of devices according to claim 5, It is characterized in that The process of efficient channel attention processing includes: The image input to the efficient channel attention is the efficient channel first image. After the efficient channel first image is processed by global average pooling and the first activation function layer in sequence, the university channel second image is obtained. The efficient channel first image and the university channel second image enter the multiplication layer to obtain the multiplication layer output image, which is the efficient channel attention output image.

7. The method for intelligently identifying ear diseases applicable to multiple types of devices according to claim 1, It is characterized in that The process of obtaining the fused image includes: Adding elements of the first feature information map and the second feature information map to obtain a first fused feature map; Extracting a first weight and a second weight of the first fused feature map; Based on the first multiplication layer, the first weight and the second feature information map are element-wise multiplied to calculate a first weight fusion feature map; Based on the second multiplication layer, the second weight and the first feature information map are element-wise multiplied to calculate a second weight fusion feature map; Based on the addition layer, the first weighted fusion feature map and the second weighted fusion feature map are element-wise added to obtain a fused image.

8. The method for intelligently identifying ear diseases applicable to multiple types of devices according to claim 3, It is characterized in that Also includes: Performing loss calculation on the second feature map to obtain a first loss value; Performing loss calculation on the third-stage image to obtain a second loss value; Performing loss calculation on the fused image to obtain a third loss value; Based on the loss function, the first loss value, the second loss value, and the third loss value are calculated to obtain a final loss value for optimizing the intelligent recognition model.

9. An intelligent ear disease recognition system applicable to multiple types of devices. It is characterized in that The system is used to implement the ear disease intelligent recognition method applicable to multiple types of devices as described in any one of claims 1 to 8, comprising: An image acquisition module is used to acquire an ear lesion image dataset, and preprocess the ear lesion image dataset to obtain a preprocessed image; A model building module, used to build an intelligent recognition model for ear diseases, wherein the intelligent recognition model includes: a downsampling module, a feature extraction module and an ear injury feature fusion module; A downsampling module, used for downsampling the preprocessed image to obtain a first feature map and a second feature map; A feature extraction module, configured to extract high-level semantic features from the first feature map to obtain a first feature information map; and extract low-level semantic features from the second feature map to obtain a second feature information map; The ear injury feature fusion module is used to fuse the first feature information map and the second feature information map to obtain a fused image; based on the fused image, intelligent identification of ear diseases can be achieved.

10. An electronic device, It is characterized in that The device comprises: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, the method for intelligently identifying ear diseases applicable to multiple types of devices as described in any one of claims 1 to 8 is implemented.

Images