An image processing method and system based on cone network

Abstract

The present invention discloses an image processing method based on a cone network, including: designing a cone network structure; performing feature extraction in the encoder part to obtain a coding vector of corresponding information; designing a decoder part structure, introducing a second decoder based on the original UNet decoder structure, further mining and utilizing the feature information extracted in the decoder, and improving the accuracy and detail recovery ability in the decoding process; fusing the features of the encoder and the first decoder through jump connections to make the decoding process more accurate and effective; training the cone network model; and evaluating the performance of the model using evaluation indicators. The present invention effectively solves the common problems of information loss and image detail blur in traditional image processing networks, and is applicable to multiple image processing fields, such as image classification, image segmentation, image registration, etc., and can greatly improve image processing efficiency and quality.

Description

Technical Field

[0001] The invention belongs to the field of image processing, and in particular relates to an image processing method and system based on a cone network. Background Art

[0002] With the rapid development of science and technology, research based on deep learning plays an increasingly important role in image processing. By simulating the operation of the human brain neural network, computers can automatically extract and learn features from large amounts of data, thereby achieving efficient processing and understanding of images. Traditional image processing methods often rely on manually designed features and algorithms, which are incapable of processing complex and changing images. Deep learning can automatically learn the feature representation in the image, thereby achieving accurate recognition and understanding of the image content.

[0003] As one of the important representatives of deep learning in the field of image processing, U-Net's structural design and performance optimization make it shine in applications such as image classification, image segmentation, and image registration. U-Net combines the structure of encoders and decoders to achieve layer-by-layer extraction and fusion of image features, thereby improving the accuracy and efficiency of image segmentation. In addition, with the continuous development of deep learning technology, U-Net is also constantly improving and optimizing to adapt to more complex and diverse image processing tasks.

[0004] However, like many deep learning models, U-Net also faces some inherent problems, which limit its performance in registration tasks to a certain extent, such as information loss, difficulty in capturing global information, and rough processing of image edges. These problems need to be further improved. Summary of the invention

[0005] The purpose of the present invention is to provide an image processing method based on a cone network model, which aims to solve some problems existing in the current U-Net network, such as information loss and difficulty in capturing global information, and to improve the image processing accuracy and performance of the network.

[0006] The technical solution of the present invention is as follows:

[0007] An image processing method based on a cone network comprises the following steps:

[0008] S1 obtains the image data set required for model training according to the needs of different image processing tasks;

[0009] S2 preprocesses the data set;

[0010] S3 divides the data set into training set and test set;

[0011] S4 establishes a cone network model according to the image processing task, and trains the cone network model through the training set;

[0012] S5 tests the cone network model trained with the training set through the test set;

[0013] S6 inputs the image data into the trained cone network model to obtain the corresponding output result.

[0014] Furthermore, the preprocessing in S2 includes resampling, normalization and image data enhancement.

[0015] Furthermore, the normalization is specifically: scaling the pixel or voxel intensity in the data set to [0, 1].

[0016] The data enhancement methods include cropping, expansion, random rotation, and flipping operations.

[0017] Furthermore, the S4 is specifically:

[0018] S41: Design of encoder for cone network model;

[0019] S42: Design bottleneck layer of cone network;

[0020] S43: Design the first decoder of the cone network,

[0021] S44: Design the second decoder of the cone network,

[0022] S45: Design an image prediction head according to a specific image processing task.

[0023] Furthermore, the S45 is specifically as follows:

[0024] S451: clarify the goal of the image processing task and determine the result required for the image task output;

[0025] S452: Calculate the loss of the output result;

[0026] S453: Back propagation loss optimization network completes training.

[0027] An image processing system based on a cone network comprises: an image acquisition unit, an image preprocessing unit and an image processing unit;

[0028] The image acquisition unit is used to acquire the image required for the image processing task;

[0029] The image preprocessing unit is used to perform a series of preprocessing operations on the acquired image data to improve the visual effect of the image;

[0030] The image processing unit includes a built cone network model for implementing high-level image processing tasks.

[0031] Furthermore, the cone network model includes an encoder, a bottleneck layer, two decoders and an image prediction head, wherein the encoder and the bottleneck layer are connected in sequence, and then the two decoders are respectively connected. At the same time, each layer of the encoder is jump-connected to the corresponding layers of the two decoders, and the corresponding layers of the first decoder and the second decoder are also jump-connected.

[0032] Furthermore, the encoder comprises four layers, and the structure of each layer consists of a maximum pooling layer and two convolution-BN-Relu activation units, and the maximum pooling layer and the two convolution-BN-Relu activation units are connected in sequence.

[0033] Furthermore, the bottleneck layer includes a maximum pooling layer, two convolution-BN-Relu activation units and a transposed convolution, and the maximum pooling layer, the two convolution-BN-Relu activation units and the transposed convolution are connected in sequence.

[0034] Furthermore, the two decoders have the same structure, comprising four layers, and the structure of each layer consists of a transposed convolution, a concat operation, and a convolution-BN-Relu activation unit, and the transposed convolution, the concat operation, and the convolution-BN-Relu activation unit are connected in sequence.

[0035] Compared with the prior art, the present invention has the following advantages:

[0036] The present invention improves the UNet network model and proposes a new type of universal image processing network model, the cone network model, which is suitable for a variety of image processing downstream tasks.

[0037] The cone network model of the present invention creatively adds a decoder on the basis of the original UNet network, forming a unique dual decoder structure. This design enables the network to perform multi-stage decoding of the extracted features compared to the original UNet encoder. In this way, the network can more effectively capture various detailed information of the image, thereby enhancing the deep understanding and processing capabilities of the image content.

[0038] The cone network model of the present invention is extremely flexible. Its encoder, decoder 1 and decoder 2 can all be combined with different network structures and parameters to adapt to different image processing tasks and data sets. This structural design makes the cone network model more advantageous in performance optimization and adaptability, providing researchers with a broad space for improvement and innovation. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The accompanying drawings generally illustrate various embodiments by way of example and not limitation, and together with the description and claims, serve to illustrate the embodiments of the invention. Where appropriate, the same reference numerals are used throughout the drawings to refer to the same or similar parts. Such embodiments are illustrative and are not intended to be exhaustive or exclusive embodiments of the present apparatus or method.

[0040] Figure 1 A schematic diagram of the process flow of the image processing method based on the cone network provided by the present invention;

[0041] Figure 2 The overall structure diagram of the cone network designed for the present invention, here taking the basic cone network as an example, that is, decoder 2 (PC) uses UNet;

[0042] Figure 3 The overall structure diagram of the cone network using the UNet improved network as decoder 2. The decoder 2 (PC) in this figure uses UNet3+;

[0043] Figure 4 The overall structure diagram of the cone network using the UNet improved network as decoder 2. The decoder 2 (PC) in this figure uses DenseUNet;

[0044] Figure 5 is a schematic flow chart of step S4;

[0045] Figure 6 is a flow chart of step S45;

[0046] Figure 7 FIG. 1 is a flow chart of the image processing system described in step S6. DETAILED DESCRIPTION

[0047] 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.

[0048] like Figure 1 As shown, an image processing method based on a cone network comprises the following steps:

[0049] Step S1: Obtain the required image data set according to the needs of different image processing tasks;

[0050] Step S2: preprocessing the data set;

[0051] Step S3: Divide the data set into a training set and a test set;

[0052] Step S4: designing a cone network according to the task and training the cone network model through training set images;

[0053] Step S5: Testing the cone network trained with the training set images by using the images of the test set;

[0054] Step S6: Input the image data into the trained cone network to obtain the corresponding output result.

[0055] For step S1: according to the needs of different image processing tasks, the required image data set is obtained.

[0056] Specifically, the data collection is to download corresponding public image data sets according to task requirements, or to collect and produce them using equipment or systems.

[0057] For step S2: preprocess the data set.

[0058] Specifically, the preprocessing methods include resampling, normalization and image data enhancement, where data enhancement includes cropping, expansion, random rotation, flipping operations, etc.

[0059] Normalization: Normalize the image and scale the pixel or voxel intensity in the dataset to [0,1]. The pixel or voxel intensity outside the interval is set to 0 and 1, respectively.

[0060] Data augmentation: Its purpose is to generate more training samples by transforming the original images, thereby enhancing the generalization ability of the model. This method can simulate various changes that may occur in real scenes, so that the model can better adapt to various complex situations.

[0061] For step S3: divide the images into training set and test set.

[0062] Specifically, this step ensures that the model can be learned on one subset (training set) and then evaluated on another independent subset (test set) to verify the generalization ability of the model. A common division ratio is 80% training set and 20% test set, but this can be adjusted according to the specific task and dataset size.

[0063] For step S4, establishing the cone network model specifically includes the following steps:

[0064] Step S41: Design the encoder structure of the cone network; Figure 2 As shown in Figure 1, the AP stage is the encoder structure. The encoder performs downsampling at each level to expand the receptive field of the model, extract the semantic features of the image, and input the features at each level to the corresponding decoders at each level through skip connections, including decoder 1 represented by PB and decoder 2 represented by PC.

[0065] When designing the encoder, taking the UNet network as an example, the nodes of each layer include two consecutive convolution-BN-Relu activation units and a maximum pooling layer. The convolution kernel size is 3×3×3, the stride is 1, the padding is 1, the pooling kernel of the maximum pooling layer is 2×2×2, the stride is 2, and the three dimensions of the input are downsampled by 2 times to expand the receptive field.

[0066] Step S42: Design the bottleneck layer structure of the cone network; design a bottleneck layer before the upsampling process to reduce the number of feature maps, which can reduce the computational complexity of the network while retaining important information. Taking the UNet network as an example, it includes a maximum pooling layer, two convolution-BN-Relu activation units and a transposed convolution, the convolution kernel size is 3×3×3, the stride is 1, the padding is 1, the pooling kernel of the maximum pooling layer is 2×2×2, the stride is 2, and the convolution kernel of the transposed convolution is 2×2×2, the stride is 2.

[0067] Step S43: Design the decoder 1 structure of the cone network; Decoder 1 part ( Figure 2 The decoder uses skip connections to fuse the low-level features in the encoder with the high-level features in the decoder, which helps improve the accuracy of the results.

[0068] Symmetrical to the encoder part, taking the UNet decoder as an example, each node includes a 2×2×2 transposed convolution operation with a stride of 2, which is then concatenated with the feature map of the encoding part, and finally passed through two 3×3×3 convolutions (using the ReLU activation function).

[0069] Step S44: Design the structure of the decoder 2 of the cone network; Decoder 2 ( Figure 2 The PC stage shown in the figure decodes the feature map decoded by decoder 1 again. By introducing a second decoder, the network can further mine and utilize the feature information extracted in the decoder to improve the accuracy and detail recovery ability of the decoding process. When using a jump connection, the feature map of the corresponding layer of the encoder and the feature map of the corresponding layer of decoder 1 are fused into decoder 2. This jump connection method can not only combine the shallow features in the encoder with the deep features in the decoder, but also enable the features obtained from decoder 1 to guide the features obtained from the encoder. Through this guidance, the deep features can guide the shallow features, making the decoding process more accurate and effective, and improving the performance of the entire network.

[0070] Taking the UNet decoder as an example, each node includes a 2×2×2 transposed convolution operation with a stride of 2, and then concatenates the feature maps with the features of the encoding part and the decoder 1 part, and finally passes through two 3×3×3 convolutions (using the ReLU activation function).

[0071] Each part of the cone network structure described above takes the structure of the UNet network as an example to form the most basic cone network model for processing three-dimensional images, that is, a UNet decoder is added again on the basis of the UNet network model.

[0072] If you need to process a two-dimensional image, the convolution dimension of the model needs to be reduced by one dimension, from 3D convolution to 2D convolution. That is, the 3×3×3 convolution becomes a 3×3 convolution, the stride and padding remain unchanged and are set to 1, the pooling kernel of the maximum pooling layer uses 2×2, the stride remains unchanged and is set to 2, and the convolution kernel of the transposed convolution uses 2×2, and the stride remains unchanged and is set to 2.

[0073] In addition, this cone network can adopt more combinations and optimization methods, as follows:

[0074] 1. Use UNet variants: You can consider using UNet variants, such as UNet++, UNet3+ (such as Figure 3 as shown) or DenseUNet (as Figure 4 As shown in Figure 2, these variants are improved and optimized based on UNet, which can further improve the performance of the model.

[0075] 2. Transformer optimization: The Transformer structure has achieved remarkable results in the field of natural language processing and can also be applied to image processing tasks. By adding the Transformer structure to the bottleneck layer, the model's ability to understand long-distance dependencies of images can be optimized, improving the accuracy of segmentation or classification.

[0076] 3.MLP mechanism or attention mechanism: Skip connection is an important part of the UNet structure, which helps the model to utilize feature information at different levels at the same time. By adding MLP (multi-layer perceptron) mechanism or attention mechanism, the effect of skip connection can be further improved, so that the model can better focus on important features.

[0077] Step S45, designing an image prediction head according to a specific image processing task includes the following steps:

[0078] Step S451: clarify the goal of the image processing task, such as image classification, image segmentation, image registration, etc., and determine the required result of the image task output.

[0079] If it is a classification task, the classification head is designed to output the probability of each category, and the channel dimension is compressed and the features are fused through convolution with a convolution kernel of 1, and then combined with global pooling to capture the global information of the entire image, so that the model can better understand the overall characteristics of the image. Finally, the features are nonlinearly transformed and classified through the fully connected layer, so that the model can output the probability of the image in each category.

[0080] If it is a segmentation task, the segmentation head is designed to output the category label of each pixel. The segmentation head uses a convolution operation with a convolution kernel of 1 to convert the number of channels to the total number of categories in semantic segmentation, and converts the feature map extracted by the previous network into pixel-level classification predictions. After completing the convolution operation, the segmentation head outputs a multi-channel feature map of the same size as the input image, where each channel corresponds to a category. Then, in order to convert these raw scores into probability distributions, the Softmax activation function is usually used. The Softmax function normalizes all category scores for each pixel to ensure that their sum is 1, thereby representing the predicted probability of each category.

[0081] For registration tasks, a deformation field φ is designed to be generated. The deformation field is usually a mapping of spatial transformations that guides how to distort and deform the moving image to align it with the fixed image. A convolution operation with a convolution kernel of 1 is used to output a deformation field with 3 channels. After the deformation field is generated, we use a spatial transformer network (STN) to distort the moving image. The STN network is a learnable module that can spatially transform the image according to the instructions of the deformation field. By combining the deformation field with the moving image, the STN network can calculate the new position of each pixel and resample and interpolate the image according to these new positions to obtain the registered image.

[0082] S452: Calculate the loss for the output result.

[0083] The trained cone network model obtained in step 3 is saved and tested. The effect of the model is evaluated using evaluation indicators based on the test results. The evaluation indicators here use the indicators required for the image processing task to be performed.

[0084] For classification tasks, the cross entropy loss function is calculated, which measures the difference between the predicted probability distribution and the true probability distribution. In image classification tasks, the cross entropy loss function is often used with the softmax activation function to convert the raw output of the model into a probability distribution.

[0085]

[0086] N represents the number of samples; K represents the number of categories; y ic Represents the one-hot encoding of the sample target value. If the sample p i If the true category is equal to c, it takes 1, otherwise it takes 0; h θ (x i ) c Represents sample x i The probability of belonging to category c. The more consistent the model's predicted probability distribution is with the true label probability distribution, the smaller the value of the cross entropy loss function is. The cross entropy loss function can be directly used in the gradient descent algorithm to update the model's weights.

[0087] If it is a segmentation task, calculate the Dice coefficient to compare the similarity of two samples. It is usually used to evaluate the similarity of image segmentation or other binarization results, and is widely used in the field of medical image segmentation. The calculation formula of DSC is:

[0088]

[0089] Among them, X and Y are the binarization results of the two samples, |X∩Y| represents the number of elements common to the two samples, and |X| and |Y| represent the total number of elements of the two samples.

[0090] If it is a registration task, calculate the image similarity between the fixed image and the registered image and deformation field smoothness Here The NCC normalized cross correlation (NCC) loss is used for calculation, which obtains the correlation between X and Y by calculating the covariance of two images X and Y divided by the standard deviation between the two images. The gradient of the displacement field u is used for calculation to ensure the physical smoothness of the generated deformation field. The formulas are as follows:

[0091]

[0092] Among them, X and Y represent an image respectively. Consistent with the calculation method of the correlation coefficient, the normalized cross-correlation is obtained by calculating the covariance of image X and image Y divided by the standard deviation of image X and the standard deviation of image Y to obtain the correlation between the two images.

[0093]

[0094] Among them, φ represents the deformation field, Ω represents the set of spatial positions of each pixel or voxel in the image, u represents the total displacement field of each point, and u(p) represents the displacement of the image pixel or voxel at the spatial position p.

[0095] Step S453, back propagation loss optimization network completes training.

[0096] Step S46: Complete the training of the cone network.

[0097] For step S5, the trained cone network model is tested using the test set:

[0098] Load the model parameters of the model trained in the previous step and input the preprocessed test set images into the image.

[0099] The results are evaluated using different evaluation criteria according to different tasks:

[0100] If it is a classification task, calculate the accuracy, precision, and recall.

[0101] The accuracy rate indicates the proportion of the number of correctly identified samples in the total number of samples. In the following formula, TN is the number of samples predicted to be negative and actually negative; TP is the number of samples predicted to be positive and actually positive; FP is the number of samples predicted to be positive and actually negative; FN is the number of samples predicted to be negative and actually positive.

[0102]

[0103] Precision indicates the proportion of true positive samples among samples identified as positive samples.

[0104]

[0105] The recall rate indicates the proportion of correctly identified positive samples in the total positive samples.

[0106]

[0107] If it is a segmentation task, calculate the Dice coefficient to compare the similarity of two samples. It is usually used to evaluate the similarity of image segmentation or other binarization results, and is widely used in the field of medical image segmentation. The calculation formula of the Dice coefficient (DSC) has been given above.

[0108] If it is a registration task, calculate the Dice coefficient (DSC) and the negative value of the Jacobian determinant The ratio is used as an evaluation indicator, and its Jacobian determinant is calculated at each point of the deformation field. When its value is between 0 and 1, it means that the volume of the point has shrunk; when its value is less than or equal to 0, it means that the image has folded at this point. That is, when the Jacobian determinant is less than 0, its smoothness can be quantified by calculating the Jacobian determinant of each image. The smaller the value, the less folding and the smoother the deformation field. The Jacobian determinant of each deformation field obtained in the test set is averaged. The smaller the value, the better the overall registration effect.

[0109] The calculation formula of Dice coefficient (DSC) is given above, and the calculation formula of Jacobi determinant is as follows. Here we take the calculation formula for three-dimensional image as an example:

[0110]

[0111] For step S6, the image data is input into the trained cone network to obtain the corresponding output result.

[0112] The present application also includes an image processing system based on a cone network, including: an image acquisition unit, an image preprocessing unit and an image processing unit. Each unit assumes a specific function and works together to complete the image processing task. Among them, the image acquisition unit is used to obtain the image required for the image processing task. Whether it is captured in real time by a camera or read from a database, network or other storage device, the unit can effectively obtain the required image resources. The image preprocessing unit is used for the unit to perform a series of preprocessing operations on the acquired image data, such as denoising, enhancement, normalization, etc., to improve the visual effect of the image and make it more suitable for subsequent cone network processing. The image processing unit contains a built cone network. This model can automatically learn and extract key features in the image through deep learning technology, thereby realizing advanced image processing tasks such as classification, segmentation and registration. In the actual operation of the system, the collected image data set must first be divided into a training set and a test set. The training set is used to train the cone network. Through a large amount of image data and corresponding labels, the network can learn how to accurately identify and process images. The test set is used to evaluate the performance of the trained model to ensure that it can maintain high accuracy when facing new image data. During the model training and testing phase, the system will continuously adjust and optimize the parameters of the cone network through a series of algorithms and optimization strategies to improve the generalization and robustness of the model. Once the model shows good performance in the test set, it can be used to process new image data. At this time, the image to be processed is input into the image processing unit, and the cone network will accurately analyze and process the image based on the knowledge learned during its training process, and finally output the required processing results.

[0113] The cone network model includes an encoder, a bottleneck layer, two decoders and an image prediction head, wherein the encoder and the bottleneck layer are connected in sequence, and then the two decoders are connected respectively. At the same time, each layer of the encoder is jump-connected to the corresponding layers of the two decoders, and the corresponding layers of the first decoder and the second decoder are also jump-connected.

[0114] The encoder comprises four layers, and the structure of each layer consists of a maximum pooling layer and two convolution-BN-Relu activation units, and the maximum pooling layer and the two convolution-BN-Relu activation units are connected in sequence.

[0115] The bottleneck layer includes a maximum pooling layer, two convolution-BN-Relu activation units and a transposed convolution, and the maximum pooling layer, the two convolution-BN-Relu activation units and the transposed convolution are connected in sequence.

[0116] The two decoders have the same structure and include four layers. The structure of each layer consists of a transposed convolution, a concat operation and a convolution-BN-Relu activation unit. The transposed convolution, the concat operation and the convolution-BN-Relu activation unit are connected in sequence.

[0117] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. An image processing method based on a cone network, characterized in that: The following steps are involved: S1 obtains the image data set required for model training according to the needs of different image processing tasks; S2 preprocesses the data set; S3 divides the data set into training set and test set; S4 establishes a cone network model according to the image processing task, and trains the cone network model through the training set; S5 tests the cone network model trained with the training set through the test set; S6 inputs the image data into the trained cone network model to obtain the corresponding output result; The preprocessing in S2 includes resampling, normalization and image data enhancement; The normalization is specifically: scaling the pixel or voxel intensity in the data set to [0, 1]; The data enhancement methods include cropping, expansion, random rotation, and flipping operations; The S4 is specifically: S41: Design of encoder for cone network model; S42: Design bottleneck layer of cone network; S43: Design the first decoder of the cone network; S44: Design the second decoder of the cone network; S45: designing an image prediction head according to a specific image processing task; The bottleneck layer includes a maximum pooling layer, two convolution-BN-Relu activation units and a transposed convolution, and the maximum pooling layer, the two convolution-BN-Relu activation units and the transposed convolution are connected in sequence; The two decoders have the same structure and include four layers. The structure of each layer consists of a transposed convolution, a concat operation and a convolution-BN-Relu activation unit. The transposed convolution, the concat operation and the convolution-BN-Relu activation unit are connected in sequence.

2. The image processing method based on cone network according to claim 1, characterized in that: The S45 is specifically: S451: clarify the goal of the image processing task and determine the result required for the image task output; S452: Calculate the loss of the output result; S453: Back propagation loss optimization network completes training.

3. An image processing system based on a cone network, characterized in that: include: Image acquisition unit, image preprocessing unit and image processing unit; The image acquisition unit is used to acquire the image required for the image processing task; The image preprocessing unit is used to perform a series of preprocessing operations on the acquired image data to improve the visual effect of the image; The preprocessing includes resampling, normalization and image data enhancement; The normalization is specifically: scaling the pixel or voxel intensity in the data set to [0, 1]; The data enhancement methods include cropping, expansion, random rotation, and flipping operations; The image processing unit includes a built cone network model for implementing high-level image processing tasks; Specifically: S41: Design of encoder for cone network model; S42: Design bottleneck layer of cone network; S43: Design the first decoder of the cone network; S44: Design the second decoder of the cone network; S45: designing an image prediction head according to a specific image processing task; The bottleneck layer includes a maximum pooling layer, two convolution-BN-Relu activation units and a transposed convolution, and the maximum pooling layer, the two convolution-BN-Relu activation units and the transposed convolution are connected in sequence; The two decoders have the same structure and include four layers. The structure of each layer consists of a transposed convolution, a concat operation and a convolution-BN-Relu activation unit. The transposed convolution, the concat operation and the convolution-BN-Relu activation unit are connected in sequence.

4. The image processing system based on cone network according to claim 3, characterized in that: The cone network model includes an encoder, a bottleneck layer, two decoders and an image prediction head, wherein the encoder and the bottleneck layer are connected in sequence, and then the two decoders are connected respectively. At the same time, each layer of the encoder is jump-connected to the corresponding layers of the two decoders, and the corresponding layers of the first decoder and the second decoder are also jump-connected.

5. The image processing system based on cone network according to claim 4, characterized in that: The encoder comprises four layers, and the structure of each layer consists of a maximum pooling layer and two convolution-BN-Relu activation units, and the maximum pooling layer and the two convolution-BN-Relu activation units are connected in sequence.

Images