CN109034249B - Convolution optimization method, device, terminal device and computer-readable storage medium based on decomposing radially symmetric convolution kernel - Google Patents

Abstract

The invention discloses a convolution optimization method, device, terminal device and computer-readable storage medium based on decomposing radially symmetric convolution kernels. The method includes: inputting an image to be recognized, and preprocessing the image to be recognized; A 1*1 convolution kernel and (m‑1)/2 1*m (m=2k+3, k∈N) convolution kernels obtained by pre-decomposing m*m radially symmetric convolution kernels respectively kernel, convolve the preprocessed image to be recognized, and obtain a 1*1 first feature map and (m‑1)/2 1*m (m=2k+3, k∈N) first feature maps Two feature maps; further convolve the second feature map to obtain a third feature map; sum the first feature map and the (third feature map to obtain a target feature map, and output the target feature map. The present invention By reducing the amount of parameters on the basis of reducing the amount of calculation of the radially symmetric convolution kernel, the purpose of optimizing the convolution is achieved.

Description

Convolution optimization method, device, terminal device and computer-readable storage medium based on decomposing radially symmetric convolution kernel

technical field

The present invention relates to the technical field of neural networks, and in particular, to a convolution optimization method, device, terminal device and computer-readable storage medium based on decomposing radially symmetric convolution kernels.

Background technique

Convolutional Neural Network (CNN) is the most commonly used neural network in the field of image processing and recognition in recent years. It has the advantages of good feature classification effect and easy processing of high-dimensional data, but convolutional neural network is prone to overfitting. Therefore, in the traditional construction of convolutional neural networks, the original training set images are often mirrored and rotated at a large angle to increase the performance of the convolutional neural network. Robustness, so that the convolutional neural network can recognize pictures from any angle; but this traditional method will increase the amount of data and increase the training time.

In the prior art, in view of the problems existing in the traditional methods, convolution kernels with radial symmetry properties are usually used, which can provide good robustness in the use of convolutional neural networks, and can reduce the occurrence of overfitting. the possibility of the phenomenon. However, the radially symmetric convolution kernel requires too much computation in the operation. The existing technical solutions for the problem of the too large amount of computation are mainly to use the convolution kernel cropping method and the multi-channel volume in the convolutional neural network. Product optimization algorithm. However, when the inventor implements the prior art solutions, it is found that these optimization algorithms may have a problem of excessively large amount of parameters in use.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is to provide a convolution optimization method, device, terminal device and computer-readable storage medium based on decomposing radially symmetric convolution kernels. The amount of parameters, so as to achieve the purpose of optimizing the convolution.

In order to solve the above technical problems, an embodiment of the present invention provides a convolution optimization method based on decomposing radially symmetric convolution kernels, suitable for execution in a computing device, including the following steps:

inputting an image to be recognized, and preprocessing the image to be recognized;

Use a 1*1 convolution kernel and (m-1)/2 1*m (m=2k+3, k∈N) convolution kernels obtained by pre-decomposing m*m radially symmetric convolution kernels respectively. kernel, convolve the preprocessed image to be recognized, and obtain a 1*1 first feature map and (m-1)/2 1*m (m=2k+3, k∈N) first feature maps Two feature maps; then use the pre-decomposition m*m radially symmetric convolution kernel to obtain (m-1)/2 convolution kernels of 1*m (m=2k+3, k∈N) one-to-one correspondence (m-1)/2 convolution kernels of m*1 (m=2k+3, k∈N), for (m-1)/2 1*m (m=2k+3, k∈N) The second feature map is convolved to obtain (m-1)/2 m*1 (m=2k+3, k∈N) third feature maps;

Sum up one 1*1 first feature map and (m-1)/2 m*1 (m=2k+3, k∈N) third feature maps to obtain the target feature map, and output the target feature map.

Further, the preprocessing of the to-be-recognized image is specifically:

According to preset parameters, randomly stretch and adjust the brightness and darkness of the image to be recognized, and add specific Gaussian noise;

Further, according to the requirements of the convolution processing, the to-be-recognized image is rotated and cut at an angle of 0-π/2.

Further, the convolution kernel matrix A of each convolution kernel satisfies the following formula:

in,

Further, each m*1 (m=2k+3, k∈N) convolution kernel and a corresponding m*1 (m=2k+3, k∈N) convolution kernel form a proportional symmetry Vector group, or ISV, specifically:

Let a _n be the parameter value, a ₁ is always equal to 1, the length of the ISV is m (m=2k+3, k∈N),

ISV=(ISV_1,ISV_2);

Among them, ISV_1 is a 1*m vector, and ISV_2 is an m*1 vector;

Further, after summing the first feature map of 1*1 and the third feature map of (m-1)/2 m*1(m=2k+3,k∈N), we get target feature map, and after outputting the target feature map, it also includes:

Determine whether the direction of the image has an influence on the recognition result;

If so, perform global average pooling on a first feature map of 1*1 and a third feature map of (m-1)/2 m*1 (m=2k+3, k∈N). The output 1+(m-1)/2 values are processed by softmax to obtain the probability of the target feature map.

Further, the described convolution optimization method based on decomposing radially symmetric convolution kernels also includes:

If not, perform spatial pyramid pooling and full-scale spatial pyramid pooling on a first feature map of 1*1 and a third feature map of (m-1)/2 m*1 (m=2k+3, k∈N). The connection layer is processed, and the output 1+(m-1)/2 values are subjected to softmax processing to obtain the judgment result of the target feature map.

An embodiment of the present invention also provides a convolution optimization device based on decomposing radially symmetric convolution kernels, including:

an input module for inputting an image to be recognized and preprocessing the image to be recognized;

The convolution optimization module is used to use a 1*1 convolution kernel and (m-1)/2 1*m(m=2k+3, k∈N) convolution kernel, convolve the preprocessed image to be recognized, and obtain a 1*1 first feature map and (m-1)/2 1*m(m=2k+3 ,k∈N) of the second feature map; and then use the pre-decomposition m*m radially symmetric convolution kernel and (m-1)/2 1*m(m=2k+3,k ∈N) Convolution kernels correspond to (m-1)/2 m*1 (m=2k+3, k∈N) convolution kernels, for (m-1)/2 1*m (m=2k The second feature map of +3,k∈N) is convolved to obtain the third feature map of (m-1)/2 m*1(m=2k+3,k∈N);

The output module is used to sum the first feature map of 1*1 and the third feature map of (m-1)/2 m*1 (m=2k+3, k∈N) to obtain the target feature map, and output the target feature map.

Further, each m*1 (m=2k+3, k∈N) convolution kernel and a corresponding m*1 (m=2k+3, k∈N) convolution kernel form a proportional symmetry Vector group, or ISV, specifically:

Let a _n be the parameter value, a ₁ is always equal to 1, the length of the ISV is m (m=2k+3, k∈N),

ISV=(ISV_1,ISV_2);

Among them, ISV_1 is a 1*m vector, and ISV_2 is an m*1 vector;

An embodiment of the present invention also provides a convolution optimization terminal device based on decomposing radially symmetric convolution kernels, comprising: a processor, a memory, and a function stored in the memory and configured to be executed by the processor A computer program, when the processor executes the computer program, the above-mentioned convolution optimization method based on decomposing radially symmetric convolution kernels is implemented.

An embodiment of the present invention further provides a computer-readable storage medium, where the computer-readable storage medium includes a stored computer program, wherein when the computer program runs, the device where the computer-readable storage medium is located is controlled to execute The above-mentioned convolution optimization method based on decomposing radially symmetric convolution kernels.

Compared with the prior art, the beneficial effects of an embodiment of the present invention are:

The convolution optimization method, device, terminal device, and computer-readable storage medium based on decomposing radially symmetric convolution kernels provided by the present invention include: inputting an image to be recognized, and preprocessing the image to be recognized; Use a 1*1 convolution kernel and (m-1)/2 1*m (m=2k+3, k∈N) convolution kernels obtained by pre-decomposing m*m radially symmetric convolution kernels respectively. kernel, convolve the preprocessed image to be recognized, and obtain a 1*1 first feature map and (m-1)/2 1*m (m=2k+3, k∈N) first feature maps Two feature maps; re-use (m-1)/2 m*1(m =2k+3,k∈N) convolution kernel, convolve the second feature map of (m-1)/2 1*m(m=2k+3,k∈N), get (m- 1)/2 m*1 (m=2k+3, k∈N) third feature maps; for 1 1*1 first feature map and (m-1)/2 m*1(m =2k+3,k∈N) and the third feature maps are summed to obtain the target feature map, and output the target feature map. The invention achieves the purpose of optimizing the convolution by reducing the amount of parameters on the basis of reducing the amount of calculation of the radially symmetric convolution kernel.

Description of drawings

1 is a schematic flowchart of a convolution optimization method based on decomposing radially symmetric convolution kernels provided by the first embodiment of the present invention;

2 is a schematic flowchart of a process of decomposing a 3*3 radially symmetric convolution check image in the first embodiment of the present invention;

3 is a schematic flow chart of the processing process of the decomposed 5*5 radially symmetric convolution check image in the first embodiment of the present invention;

4 is a schematic flow chart of a process of processing a decomposed m*m radially symmetric convolution check image in the first embodiment of the present invention;

5 is a schematic flowchart of a process for classifying cat and dog images by a neural network comprising a decomposed symmetric convolution kernel in the first embodiment of the present invention;

6 is a schematic flowchart of a process for classifying MNIST images by a neural network including a decomposed symmetric convolution kernel in the first embodiment of the present invention;

FIG. 7 is a schematic structural diagram of a convolution optimization apparatus based on decomposing radially symmetric convolution kernels provided by a second embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, so that the advantages and features of the present invention can be more easily understood by those skilled in the art. Obviously, the described The embodiments are only some of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

The first embodiment of the present invention:

See Figures 1-5.

As shown in FIG. 1 , a convolution optimization method based on decomposing a radially symmetric convolution kernel provided in this embodiment, and a convolution optimization method based on decomposing a radially symmetric convolution kernel, are suitable for execution in a computing device, including: Follow the steps below:

S101. Input an image to be recognized, and preprocess the image to be recognized;

S102, respectively using a 1*1 convolution kernel and (m-1)/2 1*m (m=2k+3, k∈N) convolution kernels obtained by decomposing m*m radially symmetric convolution kernels in advance The convolution kernel, convolves the preprocessed image to be recognized, and obtains a 1*1 first feature map and (m-1)/2 1*m (m=2k+3, k∈N) The second feature map of Corresponding (m-1)/2 m*1 (m=2k+3, k∈N) convolution kernels, for (m-1)/2 1*m (m=2k+3, k∈N) The second feature map of N) is convolved to obtain the third feature map of (m-1)/2 m*1 (m=2k+3, k∈N);

S103, summing up a first feature map of 1*1 and a third feature map of (m-1)/2 m*1 (m=2k+3, k∈N) to obtain a target feature map, And output the target feature map.

The convolution optimization method based on decomposing radially symmetric convolution kernels proposed in this embodiment enhances robustness, reduces the number of parameters, speeds up neural network fitting and prevents overfitting.

For step S101, the preprocessing of the to-be-identified image is specifically:

According to preset parameters, randomly stretch and adjust the brightness and darkness of the image to be recognized, and add specific Gaussian noise;

Further, according to the requirements of the convolution processing, the to-be-recognized image is rotated and cut at an angle of 0-π/2.

In this embodiment, it can be understood that, in order to increase the robustness, a slight Gaussian noise is added to the to-be-recognized image to be input, and small-amplitude random stretching and light and dark changes are performed. Due to the inherent advantages of the radially symmetric convolution kernel, the image to be recognized does not need to be mirrored, and the rotation step in the preprocessing only needs to rotate by an angle of 0 to π/2. Since the network only applies the convolutional neural network, there is no fully connected layer, and there is no requirement for the size. Except for the necessary cutting of the rotated image, there is no need to do any processing on the image size.

For step S102, the convolution kernel matrix A of each convolution kernel satisfies the following formula:

in,

In this embodiment, such as the following two matrices:

倍数角度时像素数值会发生变化，这里仅考虑旋转角度

A卷积处理经过θ角度旋转的图片结果与直接处理原图得到的结果相等，即处理镜像图片也会与原图产生相同的结果。Taking into account that the image is rotating non-

The pixel value will change when the angle is multiple, only the rotation angle is considered here

The result of A convolution processing the image rotated by θ angle is equal to the result obtained by directly processing the original image, that is, processing the mirror image will produce the same result as the original image.

It can be inferred from this that the above effect can be achieved by setting the convolution kernel to a matrix that satisfies A when it is initialized.

倍，

即此方法最高可以减少

的参数量。有效加快模型的收敛速度并防止过拟合。In this embodiment, the convolution kernel not only has good robustness, but also reduces the huge amount of parameters, and the parameter amount of the kernel is the same as that of the traditional convolution kernel.

times,

That is, this method can reduce the maximum

parameter amount. Effectively speed up the convergence of the model and prevent overfitting.

的参数，优化效果极为显著。This method is more effective in high-dimensional convolutional networks. In x-dimensional convolutional neural networks, the kernel can reduce at most

parameters, the optimization effect is extremely significant.

In this embodiment, further, each m*1 (m=2k+3, k∈N) convolution kernel is convolutional with a corresponding m*1 (m=2k+3, k∈N) convolution kernel The kernel forms a proportional symmetric vector group, namely ISV, specifically:

Let a _n be the parameter value, a ₁ is always equal to 1, the length of the ISV is m (m=2k+3, k∈N),

ISV=(ISV_1,ISV_2);

Among them, ISV_1 is a 1*m vector, and ISV_2 is an m*1 vector;

In this embodiment, since a large number of repeated operations are involved in the conventional convolution operation, it will consume a lot of unnecessary operation time. Therefore, this embodiment decomposes the radially symmetric convolution kernel whose length and width are 3 into A pair of ISV vectors of length 3 and a 1*1 convolution kernel are combined. The specific variables of the convolution kernel are defined as follows:

ISV_1 convolution kernel:

ISV_2 convolution kernel:

1*1 convolution kernel:

an and b in the above convolution _kernels are both parameter values.

As shown in Figure 2, the specific convolution process is as follows:

Assuming that the original image is P, first convolve P with the ISV_1 convolution kernel and the 1*1 convolution kernel respectively to obtain the feature maps P1 and P2, and then use the ISV_2 convolution kernel to convolve P1 to obtain the feature map P3, Then add P2 and P3 to get the output feature map P_output.

After mathematical operation, the convolution process in the above convolution process is equivalent to using the convolution kernel shape as

The radially symmetric convolution kernel is used for convolution, and in the above convolution process, only 7 multiplications and 5 additions are used for the same pixel, while the traditional convolution operation requires 9 multiplications and 8 additions. , which reduces a lot of computation time.

In this embodiment, as shown in FIG. 3 , preferably, if a radially symmetric convolution kernel of 5*5 is to be applied, it can be decomposed into two pairs of ISV vectors with lengths of 5 and 3 and a 1*1 Combination of convolution kernels (an _is different for each pair of ISVs).

Further, as shown in Figure 4, by extension, if a radially symmetric convolution kernel of m*m (m=2k+3, k ∈ N) is to be applied, it can be decomposed into (m-1)/2 Combine ISV vectors with lengths of 3, 5, 7, ..., m and 1*1 convolution kernels.

倍，加法的运算量为原来的

倍。In this embodiment, according to the above description, this operation method can greatly reduce the operation amount when the size of the convolution kernel is large. When the size is m*m, the operation amount of the multiplication is the same as the original.

times, the amount of operations for addition is the same as the original

times.

For step S103, further, the first feature map of 1*1 and the third feature map of (m-1)/2 m*1 (m=2k+3, k∈N) are performed. Summing to obtain the target feature map, and after outputting the target feature map, it also includes:

Determine whether the direction of the image has an influence on the recognition result;

If so, perform global average pooling on a first feature map of 1*1 and a third feature map of (m-1)/2 m*1 (m=2k+3, k∈N). The output 1+(m-1)/2 values are processed by softmax to obtain the probability of the target feature map.

In this embodiment, the convolution kernel splitting algorithm reduces the computation amount more obviously in the high-dimensional convolution kernel. As shown in Figure 5, cats and dogs are classified and recognized using the decomposed radially symmetric convolutional network.

In the construction of the neural network, an 8-layer 3*3 decomposed radially symmetric convolutional network is used at the front end of the network, and the residual layer is connected in it to prevent the gradient from disappearing.

Since the direction of the image has no effect on the recognition results, the global average pooling method is used at the end of the neural network to average pool the last two feature maps, and the two output values are processed by softmax to obtain the probability of cats and dogs.

Further, the described convolution optimization method based on decomposing radially symmetric convolution kernels also includes:

If not, perform spatial pyramid pooling and full-scale spatial pyramid pooling on a first feature map of 1*1 and a third feature map of (m-1)/2 m*1 (m=2k+3, k∈N). The connection layer is processed, and the output 1+(m-1)/2 values are subjected to softmax processing to obtain the judgment result of the target feature map.

In this example, as shown in Figure 6, a decomposed radially symmetric convolutional network is used to identify the MNIST dataset:

In the image processing stage, the above-mentioned preprocessing method is still used, but rotation is not included in the preprocessing. The main body of the neural network also uses a radially symmetric convolutional network with 8 layers of 3*3 decomposition at the front end of the network, and connects the residual layer in it to prevent the gradient from disappearing. However, due to the reduction of data, the number of layers of the neural network is correspondingly reduced.

Since the direction of the image has an impact on the recognition result, for example, the recognition of 6 and 9 will be wrong due to the uncertainty of the direction. At the end of the neural network, spatial pyramid pooling and fully connected layer with softmax output are used to judge the result of each number.

It should be noted that these are just two simple examples in this embodiment, and the network can be used in a larger and more complex identification system and achieve good results.

A convolution optimization method based on decomposing radially symmetric convolution kernels provided in this embodiment includes: inputting an image to be recognized, and preprocessing the to-be-recognized image; using pre-decomposing m*m radially symmetric convolutions respectively 1 convolution kernel of 1*1 and (m-1)/2 convolution kernels of 1*m (m=2k+3, k∈N) obtained by the accumulation kernel, perform preprocessing on the image to be recognized. Convolve to obtain a first feature map of 1*1 and (m-1)/2 second feature maps of 1*m (m=2k+3, k∈N); )/2 convolution kernels of 1*m (m=2k+3, k∈N) one-to-one correspondence (m-1)/2 volumes of m*1 (m=2k+3, k∈N) Product kernel, convolve the second feature map of (m-1)/2 1*m (m=2k+3, k ∈ N) to obtain (m-1)/2 m*1 (m= 2k+3, k∈N) of the third feature map; for a 1*1 first feature map and (m-1)/2 m*1 (m=2k+3, k∈N) the first feature map The three feature maps are summed to obtain the target feature map, and the target feature map is output. This embodiment achieves the purpose of optimizing the convolution by reducing the amount of parameters on the basis of reducing the amount of calculation of the radially symmetric convolution kernel.

The second embodiment of the present invention:

See Figure 7.

As shown in FIG. 7 , this embodiment also provides a convolution optimization device based on decomposing radially symmetric convolution kernels, including:

The input module 201 is used for inputting an image to be recognized, and preprocessing the image to be recognized.

The preprocessing of the to-be-recognized image is specifically:

According to preset parameters, randomly stretch and adjust the brightness and darkness of the image to be recognized, and add specific Gaussian noise;

Further, according to the requirements of the convolution processing, the to-be-recognized image is rotated and cut at an angle of 0-π/2.

In this embodiment, it can be understood that, in order to increase the robustness, a slight Gaussian noise is added to the to-be-recognized image to be input, and small-amplitude random stretching and light and dark changes are performed. Due to the inherent advantages of the radially symmetric convolution kernel, the image to be recognized does not need to be mirrored, and the rotation step in the preprocessing only needs to rotate by an angle of 0 to π/2. Since the network only applies the convolutional neural network, there is no fully connected layer, and there is no requirement for the size. Except for the necessary cutting of the rotated image, there is no need to do any processing on the image size.

The convolution optimization module 202 is used for respectively using a 1*1 convolution kernel and (m-1)/2 1*m (m=2k+3) obtained by decomposing m*m radially symmetric convolution kernels in advance ,k∈N) convolution kernel, convolve the preprocessed image to be recognized, and obtain a 1*1 first feature map and (m-1)/2 1*m(m=2k+ 3, k∈N) of the second feature map; then use the pre-decomposition m*m radially symmetric convolution kernel and (m-1)/2 1*m (m=2k+3, k∈N) The convolution kernels correspond to (m-1)/2 m*1 (m=2k+3, k∈N) convolution kernels, for (m-1)/2 1*m(m= The second feature map of 2k+3, k∈N) is convolved to obtain the third feature map of (m-1)/2 m*1 (m=2k+3, k∈N).

Among them, the convolution kernel matrix A of each convolution kernel satisfies the following formula:

in,

In this embodiment, such as the following two matrices:

倍数角度时像素数值会发生变化，这里仅考虑旋转角度

A卷积处理经过θ角度旋转的图片结果与直接处理原图得到的结果相等，即处理镜像图片也会与原图产生相同的结果。Take into account that the image is rotating non-

The pixel value will change when the angle is multiple, only the rotation angle is considered here

The result of A convolution processing the image rotated by θ angle is equal to the result obtained by directly processing the original image, that is, processing the mirror image will produce the same result as the original image.

It can be inferred from this that the above effect can be achieved by setting the convolution kernel to a matrix that satisfies A when it is initialized.

倍，

即此方法最高可以减少

的参数量。有效加快模型的收敛速度并防止过拟合。In this embodiment, the convolution kernel not only has good robustness, but also reduces the huge amount of parameters, and the parameter amount of the kernel is the same as that of the traditional convolution kernel.

times,

That is, this method can reduce the maximum

parameter amount. Effectively speed up the convergence of the model and prevent overfitting.

的参数，优化效果极为显著。This method is more effective in high-dimensional convolutional networks. In x-dimensional convolutional neural networks, the kernel can reduce at most

parameters, the optimization effect is extremely significant.

In this embodiment, further, each m*1 (m=2k+3, k∈N) convolution kernel is convolutional with a corresponding m*1 (m=2k+3, k∈N) convolution kernel The kernel forms a proportional symmetric vector group, namely ISV, specifically:

Let a _n be the parameter value, a ₁ is always equal to 1, the length of the ISV is m (m=2k+3, k∈N),

ISV=(ISV_1,ISV_2);

Among them, ISV_1 is a 1*m vector, and ISV_2 is an m*1 vector;

In this embodiment, since a large number of repeated operations are involved in the conventional convolution operation, it will consume a lot of unnecessary operation time. Therefore, this embodiment decomposes the radially symmetric convolution kernel whose length and width are 3 into A pair of ISV vectors of length 3 and a 1*1 convolution kernel are combined. The specific variables of the convolution kernel are defined as follows:

ISV_1 convolution kernel:

ISV_2 convolution kernel:

1*1 convolution kernel:

an and b in the above convolution _kernels are both parameter values.

As shown in Figure 2, the specific convolution process is as follows:

Assuming that the original image is P, first convolve P with the ISV_1 convolution kernel and the 1*1 convolution kernel respectively to obtain the feature maps P1 and P2, and then use the ISV_2 convolution kernel to convolve P1 to obtain the feature map P3, Then add P2 and P3 to get the output feature map P_output.

After mathematical operation, the convolution process in the above convolution process is equivalent to using the convolution kernel shape as

The radially symmetric convolution kernel is used for convolution, and in the above convolution process, only 7 multiplications and 5 additions are used for the same pixel, while the traditional convolution operation requires 9 multiplications and 8 additions. , which reduces a lot of computation time.

In this embodiment, as shown in FIG. 3 , preferably, if a radially symmetric convolution kernel of 5*5 is to be applied, it can be decomposed into two pairs of ISV vectors with lengths of 5 and 3 and a 1*1 Combination of convolution kernels (an _is different for each pair of ISVs).

Further, as shown in Figure 4, by extension, if a radially symmetric convolution kernel of m*m (m=2k+3, k ∈ N) is to be applied, it can be decomposed into (m-1)/2 Combine ISV vectors with lengths of 3, 5, 7, ..., m and 1*1 convolution kernels.

倍，加法的运算量为原来的

倍。In this embodiment, according to the above description, this operation method can greatly reduce the operation amount when the size of the convolution kernel is large. When the size is m*m, the operation amount of the multiplication is the same as the original.

times, the amount of operations for addition is the same as the original

times.

The output module 203 is used for summing a first feature map of 1*1 and a third feature map of (m-1)/2 m*1 (m=2k+3, k∈ N) to obtain target feature map, and output the target feature map.

Further, after summing the first feature map of 1*1 and the third feature map of (m-1)/2 m*1(m=2k+3,k∈N), we get target feature map, and after outputting the target feature map, it also includes:

Determine whether the direction of the image has an influence on the recognition result;

If so, perform global average pooling on a first feature map of 1*1 and a third feature map of (m-1)/2 m*1 (m=2k+3, k∈N). The output 1+(m-1)/2 values are processed by softmax to obtain the probability of the target feature map.

In this embodiment, the convolution kernel splitting algorithm reduces the computation amount more obviously in the high-dimensional convolution kernel. As shown in Figure 5, cats and dogs are classified and recognized using the decomposed radially symmetric convolutional network.

In the construction of the neural network, an 8-layer 3*3 decomposed radially symmetric convolutional network is used at the front end of the network, and the residual layer is connected in it to prevent the gradient from disappearing.

Since the direction of the image has no effect on the recognition results, the global average pooling method is used at the end of the neural network to average pool the last two feature maps, and the two output values are processed by softmax to obtain the probability of cats and dogs.

Further, the described convolution optimization method based on decomposing radially symmetric convolution kernels also includes:

If not, perform spatial pyramid pooling and full-scale spatial pyramid pooling on a first feature map of 1*1 and a third feature map of (m-1)/2 m*1 (m=2k+3, k∈N). The connection layer is processed, and the output 1+(m-1)/2 values are subjected to softmax processing to obtain the judgment result of the target feature map.

In this example, as shown in Figure 6, a decomposed radially symmetric convolutional network is used to identify the MNIST dataset:

In the image processing stage, the above-mentioned preprocessing method is still used, but rotation is not included in the preprocessing. The main body of the neural network also uses a radially symmetric convolutional network with 8 layers of 3*3 decomposition at the front end of the network, and connects the residual layer in it to prevent the gradient from disappearing. However, due to the reduction of data, the number of layers of the neural network is correspondingly reduced.

Since the direction of the image has an impact on the recognition result, for example, the recognition of 6 and 9 will be wrong due to the uncertainty of the direction. At the end of the neural network, spatial pyramid pooling and fully connected layer with softmax output are used to judge the result of each number.

It should be noted that these are just two simple examples in this embodiment, and the network can be used in a larger and more complex identification system and achieve good results.

This embodiment provides a convolution optimization device based on decomposing radially symmetric convolution kernels, by inputting an image to be recognized, and preprocessing the image to be recognized; using pre-decomposed m*m radially symmetric convolution respectively 1 convolution kernel of 1*1 and (m-1)/2 convolution kernels of 1*m (m=2k+3, k∈N) obtained by the kernel, the preprocessed image to be recognized is rolled product, to obtain a first feature map of 1*1 and (m-1)/2 second feature maps of 1*m (m=2k+3, k∈N); then use and (m-1) /2 convolution kernels of 1*m (m=2k+3, k∈N) one-to-one correspondence (m-1)/2 convolutions of m*1 (m=2k+3, k∈N) kernel, convolve the second feature map of (m-1)/2 1*m (m=2k+3, k ∈ N) to get (m-1)/2 m*1 (m=2k +3,k∈N) the third feature map; for 1 1*1 first feature map and (m-1)/2 m*1(m=2k+3,k∈N) the third feature map The feature maps are summed to obtain a target feature map, and the target feature map is output. This embodiment achieves the purpose of optimizing the convolution by reducing the amount of parameters on the basis of reducing the amount of calculation of the radially symmetric convolution kernel.

An embodiment of the present invention also provides a convolution optimization terminal device based on decomposing radially symmetric convolution kernels, comprising: a processor, a memory, and a function stored in the memory and configured to be executed by the processor A computer program, when the processor executes the computer program, the above-mentioned convolution optimization method based on decomposing radially symmetric convolution kernels is implemented.

An embodiment of the present invention further provides a computer-readable storage medium, where the computer-readable storage medium includes a stored computer program, wherein when the computer program runs, the device where the computer-readable storage medium is located is controlled to execute The above-mentioned convolution optimization method based on decomposing radially symmetric convolution kernels.

Compared with the prior art, the beneficial effects of an embodiment of the present invention are:

The convolution optimization method, device, terminal device and computer-readable storage medium based on decomposing radially symmetric convolution kernels provided by the present invention include: inputting an image to be recognized, and preprocessing the image to be recognized; Use a 1*1 convolution kernel and (m-1)/2 1*m (m=2k+3, k∈N) convolution kernels obtained by pre-decomposing m*m radially symmetric convolution kernels respectively. kernel, convolve the preprocessed image to be recognized, and obtain a 1*1 first feature map and (m-1)/2 1*m (m=2k+3, k∈N) first feature maps Two feature maps; re-use (m-1)/2 m*1(m =2k+3,k∈N) convolution kernel, convolve the second feature map of (m-1)/2 1*m(m=2k+3,k∈N), get (m- 1)/2 m*1 (m=2k+3, k∈N) third feature maps; for 1 1*1 first feature map and (m-1)/2 m*1(m =2k+3,k∈N) and the third feature maps are summed to obtain the target feature map, and output the target feature map. The invention achieves the purpose of optimizing the convolution by reducing the amount of parameters on the basis of reducing the amount of calculation of the radially symmetric convolution kernel

The above are the preferred embodiments of the present invention. It should be pointed out that for those skilled in the art, without departing from the principles of the present invention, several improvements and modifications can be made, and these improvements and modifications may also be regarded as It is the protection scope of the present invention.

Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be implemented by instructing relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. During execution, the processes of the embodiments of the above-mentioned methods may be included. The storage medium may be a magnetic disk, an optical disk, a read-only memory (Read-Only Memory, ROM), or a random access memory (Random Access Memory, RAM) or the like.

Claims (7)

1. a convolution optimization method based on decomposing radially symmetric convolution kernel, is suitable for carrying out in computing equipment, it is characterized in that, comprise the steps: Inputting the image to be recognized, and preprocessing the image to be recognized, including: randomly stretching and adjusting the brightness and darkness of the image to be recognized according to preset parameters, and adding specific Gaussian noise; according to the requirements of convolution processing , rotate and cut the to-be-recognized image at an angle of 0 to π/2; Using a 1*1 convolution kernel and (m-1)/2 1*m convolution kernels obtained by decomposing m*m radially symmetric convolution kernels in advance, the preprocessed images to be recognized are processed. Convolve to obtain a 1*1 first feature map and (m-1)/2 1*m second feature maps; then use the pre-decomposed m*m radially symmetric convolution kernel to obtain and (m -1)/2 1*m convolution kernels one-to-one (m-1)/2 m*1 convolution kernels, for (m-1)/2 1*m second feature maps Perform convolution to obtain (m-1)/2 m*1 third feature maps; Summing a first feature map of 1*1 and a third feature map of (m-1)/2 m*1 to obtain a target feature map, and output the target feature map; where m=2k+ 3, k∈N. 2. the convolution optimization method based on decomposing radially symmetric convolution kernel according to claim 1, is characterized in that, the convolution kernel matrix of each convolution kernel satisfies following formula:

in,

3. The convolution optimization method based on decomposing radially symmetric convolution kernels according to claim 1, characterized in that, in the pair of 1 first feature map of 1*1 and (m-1)/2 The third feature map of m*1 is summed to obtain the target feature map, and after outputting the target feature map, it also includes: Determine whether the direction of the image has an influence on the recognition result; If so, perform global average pooling on one 1*1 first feature map and (m-1)/2 m*1 third feature maps, and combine the output 1+(m-1)/2 Each value is subjected to softmax processing to obtain the probability of the target feature map. 4. the convolution optimization method based on decomposing radially symmetric convolution kernel according to claim 3, is characterized in that, also comprises: If not, perform spatial pyramid pooling and fully connected layer processing on one 1*1 first feature map and (m-1)/2 m*1 third feature maps, and convert the output 1+( m-1)/2 values are processed by softmax to obtain the judgment result of the target feature map. 5. a convolution optimization device based on decomposing radially symmetric convolution kernel, is characterized in that, comprises: The input module is used for inputting the to-be-recognized image and preprocessing the to-be-recognized image, including: randomly stretching and shading the to-be-recognized image according to preset parameters, and adding specific Gaussian noise; According to the requirements of convolution processing, the image to be recognized is rotated and cut at an angle of 0 to π/2; The convolution optimization module is used to use a 1*1 convolution kernel and (m-1)/2 1*m convolution kernels obtained by decomposing m*m radially symmetric convolution kernels in advance. The preprocessed image to be recognized is convolved to obtain a 1*1 first feature map and (m-1)/2 1*m second feature maps; then use the pre-decomposed m*m radial symmetry volume The product kernel obtained corresponds to (m-1)/2 1*m convolution kernels one-to-one (m-1)/2 m*1 convolution kernels, for (m-1)/2 1 The second feature map of *m is convolved to obtain (m-1)/2 third feature maps of m*1; The output module is used to sum the first feature map of 1*1 and the third feature map of (m-1)/2 m*1 to obtain the target feature map, and output the target feature map; Where m=2k+3, k∈N. 6. a convolution optimization terminal device based on decomposing radially symmetric convolution kernel, is characterized in that, comprises: processor, memory and the computer program that is stored in described memory and is configured to be executed by described processor, When the processor executes the computer program, the method for convolution optimization based on decomposing radially symmetric convolution kernels according to any one of claims 1 to 4 is implemented. 7. A computer-readable storage medium, characterized in that the computer-readable storage medium comprises a stored computer program, wherein, when the computer program is run, the device where the computer-readable storage medium is located is controlled to perform as claimed in the claims The convolution optimization method based on decomposing radially symmetric convolution kernels according to any one of 1 to 4.

Images