WO2023106870A1 - Re-structuralized convolutional neural network system using cmp and operation method thereof - Google Patents

Abstract

A re-structuralized convolutional neural network system using CMP according to the disclosed present invention comprises: a convolution layer that performs a convolution operation of an input image through a filter and outputs a feature map as a result of the operation; a pooling layer that performs sub sampling of the output feature map to reduce the size of the feature map; and a fully connected layer that flattens the feature map having the reduced size into one-dimensional array, adjusts a parameter to predict a label, proceeds to learning, and outputs a result. The pooling layer (200a, 200b) comprises a two-stage structure having: a conditional min pooling (CMP) unit (210) that, when 0 does not exist in a window, extracts a minimum value as a feature value by using min pooling and, when 0 exists in a window, does not extract 0 as a feature value, identifies the number of pixels including 0, and then configures a restriction for the ratio of the identified 0 according to an allowable value (0.25-1); and a max pooling unit (220) that extracts a maximum value from a window as a feature value.

Description

Restructured convolutional neural network system using CMP and its operation method

The present invention relates to a convolutional neural network system, and more particularly, to a restructured convolutional neural network system using Conditional Min Pooling (CMP) and an operation method thereof.

Among various studies using artificial intelligence, especially in the field of computer vision, it has become easy to acquire high-quality data due to the development of smartphone cameras and the spread of SNS (Social Networking Service). Since computer vision technology plays the same role as human eyes in computers, researches are being actively conducted in various fields such as unmanned operation of cars or drones, real-time behavior analysis, and image restoration.

However, computer vision technology has a problem in that it recognizes that there is an object even if there is no object to be detected or cannot detect it even though there is an object to be detected. Accurate analysis of the images must be performed as a necessity since such false detections can lead to accidents or conclusions in a completely different direction.

In the research direction of computer vision technology, in the past, a method of extracting and utilizing features intended by humans was used for accurate image analysis. Currently, it is changing to Deep Learning, which learns by finding important patterns and rules on its own in large-scale data. Most of the deep learning used in computer vision technology is being conducted with studies based on convolutional neural networks (CNNs).

The initial model of CNN took a long time to process a high amount of computation with the hardware performance at the time of development, and did not receive attention because it did not show remarkable performance. In addition to the high level of performance, it has attracted the attention of many researchers, and active research is being conducted for various purposes.

In particular, since changing the convolutional structure has the most direct effect on feature map extraction, most of the main studies on CNN are conducted with the changing of the convolutional structure.

In the case of pooling, it performs key functions such as reducing the amount of computation and preventing overfitting by reducing the size of feature maps, but changing the pooling structure does not have as much effect on improving CNN performance as changing the convolution structure. Little research is done on However, pooling in CNN has various problems depending on the technique used.

Common pooling techniques include Min Pooling, Average Pooling, and Max Pooling.

Min Pooling extracts the minimum value as a feature value, so it is rarely used because there is a problem that the feature itself disappears or noise can be detected as a feature in a specific image. Average Pooling calculates the average value of the two feature values when a positive feature and a negative feature exist in the same space, and the two features cancel each other out, which can lead to a situation where no feature remains. In addition, a down-scale weighting effect in which strong features are reduced may occur depending on the type of activation function. Max Pooling takes only strong features among the features arranged in space, so sophisticated features disappear. Features extracted through max pooling have problems in that accuracy is reduced or overfitting is easy in discriminating similar objects.

Stochastic pooling is being studied as a way to solve the problems of these general pooling techniques. Stochastic Pooling is a technique that utilizes probability, and like existing pooling techniques, it uses a window method. Stochastic Pooling calculates the probability for each pixel by normalizing the pixel values by dividing the entire pixel value for each pixel in the window. Then, a pixel to be extracted as a feature value is randomly selected according to the probability of the pixel. As a result, the more overlapping values, the higher the probability of being extracted as a feature value, and the higher the probability of extracting a meaningful value.

However, Stochastic Pooling tried to solve the pooling problem through probability, but when using Stochastic Pooling in practice, there is a problem with poor performance, so additional studies are being conducted to apply it.

The present invention was made in view of the above points, and provides conditional min pooling (hereinafter referred to as 'CMP') based on statistics rather than probability-based like stochastic pooling to solve the problems of existing pooling techniques. There is a purpose.

In addition, another purpose is to provide a restructured convolutional neural network system to efficiently utilize CMP.

The restructured convolutional neural network system using CMP according to the present invention for achieving the above object is a convolution layer that processes an input image through a filter and outputs a feature map as a result of the operation. And, a pooling layer that reduces the size of the feature map through sub-sampling of the output feature map and a parameter to predict the label by unpacking the size-reduced feature map into a one-dimensional array It includes a Fully Connected Layer that proceeds with learning by adjusting variables and outputs the result. Here, the pooling layer extracts the minimum value as a feature value as in Min Pooling if 0 does not exist in the window, and if 0 exists in the window, 0 is not extracted as a feature value and the number of pixels containing 0 is checked, The confirmed ratio of 0 includes a two-stage structure of a Conditional Min Pooling (CMP) unit that sets constraints according to allowable values (0.25 to 1) and a Max Pooling unit that extracts the maximum value within a window as a feature value.

The allowable value of the CMP unit may be 0.25, 0.5, 0.75, or 1.

The pooling layer converts the input image into a channel through convolution with a size of 1 × 1 for the two feature maps processed through two-stage pooling by the CMP unit and the Max Pooling unit after convolution operation is performed by the convolution layer. A first pooling layer for reducing and combining into one feature map, and after the feature maps combined through the first pooling layer repeatedly perform convolution operations on the plurality of convolution layers, before the precombination layer A second pooling layer may be included to extract two feature maps processed through two-stage pooling by the CMP unit and Max Pooling, combine them into one feature map, and send them to the pre-combination layer.

Meanwhile, a method of operating a restructured convolutional neural network system using CMP according to an embodiment of the present invention includes: a) extracting a feature map by performing a convolution operation on an input image through a convolution layer; b) The feature maps extracted in step a) are reduced through the first pooling layer, and the two feature maps processed through two-stage pooling by the CMP unit and the Max Pooling unit are convolved with a size of 1 × 1 to reduce the channel. and combining them into one feature map; c) extracting a feature map by repeatedly performing a convolution operation through a plurality of convolution layers in which the feature map combined in step b) is set; d) extracting the feature maps extracted in step c) through a second pooling layer and combining the two feature maps into one feature map after extracting two feature maps performed through two-stage pooling by the CMP unit and Max Pooling; and, e) predicting and outputting the result of the final neural network through the precombined layer of the feature map combined in step d).

According to the present invention, by proposing a CMP structure in the pooling layer, there is an effect of finding detailed feature information by solving the feature offset overfitting problem that occurs in conventional pooling, and a CNN restructured to efficiently utilize such CMP By proposing the system, more precise and diverse feature information can be found, resulting in an excellent effect of increasing the performance accuracy of the neural network model and lowering the error rate.

1 is a diagram showing the structure of a conventional general convolutional neural network system;

2 is a diagram for explaining the operation process of the convolution layer of FIG. 1;

3 is a diagram for explaining the calculation process of the pooling layer of FIG. 1;

4 is a diagram for explaining the calculation process of the precombined layer of FIG. 1;

5 is a diagram showing the structure of a restructured convolutional neural network system according to an embodiment of the present invention;

6 is a diagram for explaining the calculation process of the CMP unit of FIG. 1;

7 is a diagram for explaining the operation of a restructured convolutional neural network system according to an embodiment of the present invention;

8 is a diagram showing the final performance results of Pooling using Clatech 101 data;

9 is a diagram showing the final performance results of Pooling using crawling data;

10 is a diagram showing final performance results of a neural network model using Caltech 101 data;

11 is a diagram showing final performance results of a neural network model using crawling data.

The above objects, features and other advantages of the present invention will become more apparent by describing preferred embodiments of the present invention in detail with reference to the accompanying drawings.

First, for convenience of understanding, a conventional general convolutional neural network system will be described prior to describing a restructured convolutional neural network system using CMP according to an embodiment of the present invention.

1 shows the structure of a conventional convolutional neural network system.

As shown, the learning method of a general convolutional neural network (CNN) is based on supervised learning that provides labels for images and data. The input image is subjected to convolution operation through a filter in the convolution layer, and a feature map is output as a result of the operation. The feature map obtained in this way reduces the size of the feature map through a pooling layer. The size-reduced feature map is unfolded in one dimension, and training is performed while adjusting parameters so that the label can be predicted through a fully connected layer.

A convolution layer performs a convolution operation using a filter to extract features of an image. That is, the convolution layer extracts the features of the image and creates a feature map. 2 is a matrix-type filter having a height and a width as an operation method of a convolution layer. The window is moved at regular intervals from the pixels of the input image, and the elements corresponding to the pixels of the filter and the image are multiplied. After the operation, a single multiplication-accumulation (Fused Multiply-Add) is performed to obtain the total sum. As a result of repeating the calculation operation, a feature map of the input image is created. The size of the feature map created in this way is determined according to the size of the filter. Since different features are extracted from the feature map depending on the type of filter, multiple channels are retained by using multiple filters in the feature map.

The pooling layer is a layer that receives a feature map after calculating the convolution layer and performs a calculation (sub sampling) to reduce the spatial size of the feature map in the horizontal and vertical directions. Unlike the convolution layer, the pooling layer does not use a filter, so the number of channels does not change, and the operation method of the pooling layer is as shown in FIG. 3 . The window of the feature map moves at regular intervals, and pixel values within the window are extracted as one pixel value according to specific conditions. With this operation, the size of the feature map is reduced by reducing the number of pixels. During the pooling process, there are two methods for extracting pixel values within a window as a single pixel value. First, Max Pooling, which takes the largest value within the window, and Second, Average Pooling, which takes the average of all values within the window, is mainly applied. Since the size of the feature map that has gone through pooling is reduced, the amount of computation is reduced, and information is reduced as much as the size is reduced, so there is an effect of preventing overfitting.

The Fully Connected Layer is the last layer of the CNN structure that classifies labels by analyzing the feature map obtained by passing the image through the convolution layer and the pooling layer. The operation method of the Fully Connected Layer is shown in FIG. The structure of the fully combined layer has the form of a hidden layer and an output layer of a multi-layer perceptron, and the operation method is to unpack the finally selected feature map into a one-dimensional array and transmit it to the nodes. The transmitted data are used as computational values in the node to train the weights. Then, to proceed with prediction, Softmax is used as an activation function so that the sum of predicted values for each label becomes 1. The result with the highest probability is determined as the final prediction label, and the final neural network result is output.

Hereinafter, a restructured convolutional neural network system using CMP according to an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

5 shows the structure of a restructured convolutional neural network system using Conditional Min Pooling (CMP) according to an embodiment of the present invention.

The restructured convolutional neural network system 10 according to an embodiment of the present invention includes convolution layers 100a to 100d, pooling layers 200a and 200b, and a fully coupled layer 300, and includes a convolution layer and The fully coupled layer performs the same structure and operation as that of the conventional synthetic hole neural network.

Therefore, detailed descriptions of the convolutional layer and the fully coupled layer will be omitted, and the pooling layer will be described in detail below.

Referring to FIG. 5 , the pooling layers 200a and 200b according to an embodiment of the present invention are composed of two stages of a Conditional Min Pooling (CMP) unit 210 and a Max Pooling unit 220. As described above, the Max Pooling unit 220 extracts the maximum value within the window as a feature value, and since there is no difference from the conventional pooling method, the CMP unit 210 will be described in detail below.

CMP is designed based on Min Pooling. If 0 does not exist in the window, CMP extracts the minimum value as a feature value like Min Pooling, and if 0 exists in the window, it does not extract 0 as a feature value and pixels containing 0 After checking the number of , the ratio of confirmed 0 sets the constraint according to the allowable value (0 to 1).

Existing Min Pooling extracts the minimum value among pixel values within a window, so if there is even one 0 among pixel values within a window, the feature value is extracted as 0. This operation has the problem that all remaining features are deleted. Accordingly, the CMP proposed in the present invention can solve the problem of Min Pooling and the overfitting problem by statistically restricting the operation process of Min Pooling.

The basic operation method of CMP extracts the minimum value as a feature value in the same way as Min Pooling if 0 does not exist within the window, but if 0 exists within the window, unlike the existing Min Pooling, 0 is not extracted as a feature value and pixels containing 0 are extracted. check the number of

The percentage of zeros identified sets the constraints according to the allowable value (0.25 to 1). If the allowable value is 0, CMP operates the same as Min Pooling, so this case is excluded. Hereinafter, cases when the allowable values are 0.25, 0.5, 0.75, and 1 will be described.

If the allowable value of 0 is 0.25, 0 is extracted as a feature value when the statistical value of 0 within the window is greater than 1/4, and the minimum value excluding 0 is taken as the feature value when the statistical value is less than 1/4.

And, if the allowable value of 0 is 0.5, 0 is extracted as a feature value when the statistical value of 0 is more than half within the window, and the minimum value excluding 0 is taken as the feature value when the statistical value of 0 is less than half.

And, if the allowable value of 0 is 0.75, 0 is extracted as a feature value when the statistical value of 0 within the window is 3/4 or more, and the minimum value excluding 0 is taken as the feature value when the statistical value of 0 is less than 3/5.

Finally, if the allowable value of 0 is 1, the minimum value excluding 0 is taken as a feature value, and only when the entire window value is 0, 0 is taken as a feature value.

6 shows an operating method of the CMP unit 210 according to an embodiment of the present invention, when the window size is 2×2 and the stride is 2, the allowable value of 0 is 0.25, 0.5, 0.75, 1 As an operation method in one case, it can be confirmed that the feature map configuration changes depending on the allowable value of 0.

In this way, according to the restructured convolutional neural network structure according to the present invention, the convolution layer applies the existing convolution layer and restructures the configuration of the pooling layer. The structure of the restructured pooling layer consists of the above-described CMP unit 210 and Max Pooling unit 220, so that two pooling techniques can be used, and different pooling techniques are applied to each. As such, the two feature maps processed through two-step pooling have feature values more than twice that of the existing feature maps.

On the other hand, feature maps inevitably suffer data loss not only in the convolution process but also in the pooling process. In previous studies, many studies have been conducted to reduce data loss, but most studies are conducted based on convolutional layers with large performance changes. Pooling can extract different types of feature maps depending on the pooling technique, although not as much as the convolution layer. In addition, since the feature map that has undergone the convolution process necessarily undergoes the pooling process to reduce the amount of computation and prevent overfitting, it is essential to improve the structure of the CNN according to the pooling process.

Therefore, in consideration of this, the present invention improves the structure of the pooling layer. Referring back to FIG. 5, the pooling layer according to the present invention includes a first pooling layer 200a and a second pooling layer 200b.

The first pooling layer 200a converts the two feature maps processed through two-stage pooling by the CMP unit 210 and the Max Pooling unit 220 after calculation processing of the input image by the convolution layer 100a into 1 Channels are reduced through ×1-sized convolution, non-linearity is increased, and then combined into one feature map (filter concatenation).

The feature maps combined through the first pooling layer 200a are repeatedly subjected to convolution operations by the plurality of convolution layers 100b, 100c, and 100d.

The second pooling layer 200b extracts two feature maps that have been performed through two-step pooling by the CMP unit 210 and the Max Pooling unit 220 before the precombining layer 300 and converts them into one feature map. It is combined ((Filter concatenation).

The operation of the restructured synthetic hole neural network using CMP according to the present invention having such a structure will be described with reference to FIG. 7 .

First, the preprocessed image is input (Input Image) (S410). After the image is input, the channels of the image are separated in RGB form.

Thereafter, a feature map is extracted by performing a convolution operation and an activation function on the input image through a convolution layer (S420), where ReLU is used as an activation function.

Then, the extracted feature maps are processed through two-stage pooling (CMP, Max Poolin) by the CMP unit 410 and the Max Pooling unit 420 through the first pooling layer 200a. Channels are reduced through convolution of a size of ×1 and combined into one feature map (Filter Concatenation) (S430).

Thereafter, a feature map is extracted by repeatedly performing a convolution operation and an activation function through a plurality of convolution layers set for the combined feature map (S440).

After that, before the pre-combination layer, the extracted feature map is pooled again through the second pooling layer 200b, that is, 2-stage pooling by the CMP unit 410 and Max Pooling 420 ( After extracting the two feature maps processed through CMP, Max Poolin), they are merged into one feature map ((Filter Concatenation) (S450). Here, as in step S430, 1 1 × 1 size convolution is used. will not be

Finally, the result of the final neural network is predicted and output through the fully connected layer of the feature map combined in step S450.

Implementation and performance evaluation

Hereinafter, data collected for performance evaluation of the restructured convolutional neural network using CMP according to the present invention will be used and compared with existing studies.

1. System implementation environment

The design and implementation environment of the restructured convolutional neural network proposed in the present invention is developed and performance evaluated in the development environment shown in Table 1 below.

classification Detail OS Windows 10 CPU Intel Core i7-9700 RAM 32GB GPU GeForce RTX 2080 Super Language Python 3.6 IDE PyCharm Community 2020.1.2 Library Tensorflow1.14.0, Keras2.31

2. Performance evaluation method

Accuracy rate is used as a performance evaluation and factor to compare the performance of existing models and the model proposed in the present invention. The accuracy measurement method for evaluation in the present invention is the ratio of correctly predicted data to the total data and is as shown in Equation 1 below.

3. Training and testing data

A large amount of data is required to train the model proposed in the present invention. Methods to collect large amounts of data include the use of public data and the collection of data using crawling. In the case of public data, although the data is qualitatively reliable, there is a problem in that it is difficult to collect a data set that meets the conditions to be utilized and has a large amount of data. On the other hand, crawling can collect large amounts of data. However, there is a problem in that it is difficult to obtain accurate data due to duplication of data or the inclusion of two or more category objects in one image, thereby reducing reliability. Therefore, in this specification, performance evaluation is performed by selectively utilizing public data and data collected through crawling.

First, Caltech 101 is public data provided by the University of California, and is a total of 9,146 image data. It is a data set consisting of 101 different entities (categories), such as airplanes, ants, cameras, and chairs. The average size of each image is composed of about 300 × 200 pixels, and the amount of data per category is composed of a large difference from a minimum of 31 to a maximum of 800. Therefore, among the 12 categories with more than 100 images, the data to be used for model learning proceeds with 7 categories excluding similar or black-and-white images and categories of data collected by crawling. Table 2 shows the amount of data by category of the selected Caltech 101 data.

category amount of data average image size image average capacity Airplanes 800 402×158 10 KB Motorbikes 798 263×165 9KB Faces 435 504×333 28 KB Watch 239 292×230 14 KB Leopards 200 182×138 7KB Bonsai 128 263×281 17 KB Chandelier 107 269×274 15 KB

Next, Python's Beautifulsoup and ChromeDriver libraries are used to crawl image data, and Google image search is used to collect data. In addition, in order to increase the amount of image data collection, image data is collected through Google Image Search in Korea and Google Image Search in the United States. The data collected using crawling is inspected for image size, redundancy, and error data through the preprocessing module. Through this, data reliability is increased and a data set for model learning and testing is built through image size conversion and shape change. Table 3 classifies the data collected through crawling by category. It consists of a total of six categories: bird, boat, car, cat, dog, and rabbit, and the quantity is the amount of data before performing the Data Argument.

category amount of data average image size image average capacity bird 676 349×254 26 KB boat 557 407×284 39 KB car 702 566×357 82 KB cat 786 730×553 108 KB dog 663 696×536 122KB rabbit 500 403×302 49 KB

4. Performance evaluation of the model proposed in the present invention

Hereinafter, the performance evaluation of the CMP proposed in the present invention and the existing pooling technique is performed, and then the performance of the restructured CNN and the performance of existing algorithms are compared and evaluated. Table 4 is the structural form of the model for performance comparison evaluation of the CNN structure proposed in the present invention.

Model dataset image size network depth number of labels AlexNet Caltech 101 64x64 8 7 ResNet 64x64 56 7 DenseNet 64x64 121 7 Study of Propose 64x64 5 7 AlexNet crawl data 64x64 8 6 ResNet 64x64 56 6 DenseNet 64x64 121 6 Study of Propose 64x64 5 6

1) CMP performance evaluation

In the first performance evaluation, a pooling performance test is conducted using a part of Caltech 101 data and data collected through crawling individually. The model for performance evaluation receives an input in the size of 64x64 and consists of a structure with 2 convolution layers and 2 pooling layers. A total of 100 epochs of learning is performed by changing only the pooling method under the same conditions. Since it is advantageous for feature extraction to use Max Pooling together rather than using CMP consecutively, CMP is used for the first pooling layer and Max Pooling is used for the last pooling layer. All the rest are composed of Max Pooling or Average Pooling. In the pooling performance test, 74% of the total data was classified as training data, 16% was classified as test data during the epoch, and 10% was used as the final performance test.

Figure 8 shows the final performance results of Pooling using Clatech 101 data, the results of testing using 10% of the total data as the final accuracy and loss values of existing Pooling and Pooling (CMP, Conditional) proposed in the present invention. am. As for accuracy, Average Pooling shows the lowest value. Max Pooling and the proposed Pooling (CMP, Conditional) show similar performance results, but looking at the loss value, Max Pooling shows 0.1021% and the proposed Pooling method (CMP) shows 0.0817%. As a result of objective performance, it can be confirmed that the combination using CMP shows higher performance than the single Max Pooling combination.

Figure 9 shows the final performance results of pooling using crawling data. Looking at the overall performance, the proposed pooling (CMP, conditional) has the highest accuracy rate of 0.81 and the lowest loss rate of 0.23902. As a result of performance evaluation, it shows the highest performance among the three pooling techniques. After that, for the rest of the pooling, unlike Caltech 101, Average Pooling shows higher performance than Max Pooling.

As described above, the test results using Caltech 101 data confirmed that the accuracy rate of the pooling technique (CMP) proposed in the present invention was 0.9928%, which was 0.16 to 0.52% higher than the existing pooling technique. The loss rate was 0.0817%, which was confirmed to be 19.98~28.71% lower than the existing technique. In the test using the collected images (crawling data), the accuracy rate was 0.81%, which improved performance by 1.36~2.56% compared to the existing method, and the loss rate was 2.3902%, which was confirmed to decrease by 9.22~13.28%.

2) Restructured Convolutional Neural Network Performance Evaluation

For efficient use of CMP, the restructured convolutional neural network according to the present invention uses Caltech 101 and crawling data in the same way as the pooling performance evaluation described above, but expands the amount of existing data by 10 times with data augmentation for reliable performance evaluation. did In addition, the performance evaluation was conducted by expanding the number of times of learning from 100 times to 500 times.

10 is a final performance result of a neural network model using Caltech data. The structure proposed in the present invention has the highest accuracy rate of 0.9844%, but the error rate of AlexNet is the lowest at 0.0491%. In the case of ResNet, there is no significant difference in the performance of the structure, accuracy, and error rate proposed in the present invention, and ResNet shows better performance in some parts of the learning process, making it difficult to judge superiority or inferiority in model performance.

11 is a final performance result of a neural network model using crawling data. The accuracy rate of DenseNet is the highest at 0.8686%, and the model proposed in the present invention shows the second highest performance at 0.8473%. As for the error rate, AlexNet has the lowest error rate of 0.8454%, and the model proposed in the present invention has the second lowest error rate of 2.306%.

As described above, in the test using Caltech 101 data, the structure proposed in the present invention confirmed a performance improvement of 0.38 to 2.69% with the highest accuracy rate in the final test. As for the error rate, AlexNet showed the lowest error rate at 0.0491%, and the proposed structure ranked second with 0.2393%. Meanwhile, in the performance test using collected images (crawling data), DenseNet ranked first with 0.8686%, and the proposed structure ranked second with 0.8473%. As for the error rate, AlexNet took first place with 1.0769%, and the proposed structure took second place with 2.306%. As a result of the test, the proposed structure did not show the highest performance, but considering the fact that there was no change in the convolution structure, it is judged that it showed sufficiently remarkable performance.

Although preferred embodiments of the present invention have been described above, the present invention is not limited to the specific embodiments described above. That is, those skilled in the art to which the present invention pertains can make many changes and modifications to the present invention without departing from the spirit and scope of the appended claims, and all such appropriate changes and modifications Equivalents should also be considered as falling within the scope of this invention.

The present invention relates to a restructured convolutional neural network system using Conditional Min Pooling (CMP) used in the field of deep learning in which patterns and rules are discovered and learned by itself from large-scale image data, and an operating method thereof, which has industrial applicability. .

Claims (4)

A convolution layer processes an input image through a filter and outputs a feature map as a result of the operation, and the size of the feature map through sub sampling of the output feature map. ) and a Fully Connected Layer that solves the size-reduced feature map into a one-dimensional array, adjusts the parameters to predict the label, proceeds with learning, and outputs the result. In the synthetic hole neural network system, The pooling layer, If 0 does not exist in the window, the minimum value is extracted as a feature value as in Min Pooling. A CMP (Conditional Min Pooling) unit to set constraints according to the allowable value (0.25 to 1); A restructured convolutional neural network system using CMP, characterized in that it includes a two-stage structure of a Max Pooling unit that extracts the maximum value within the window as a feature value. According to claim 1, Restructured convolutional neural network system using CMP, characterized in that the allowable value of the CMP unit is 0.25, 0.5, 0.75, 1. According to claim 1, The pooling layer, After convolution operation processing of the input image by the convolution layer, the two feature maps processed through two-stage pooling by the CMP unit and the Max Pooling unit are reduced through convolution with a size of 1 × 1 to reduce the channel and generate one feature. A first pooling layer combining into a map; After the feature maps combined through the first pooling layer repeatedly perform convolution operations on the plurality of convolution layers, two stages of pooling performed by the CMP unit and Max Pooling before the precombination layer are performed. A restructured convolutional neural network system using CMP, characterized in that it comprises a second pooling layer for extracting two feature maps and then combining them into one feature map and sending them to the precombined layer. A method of operating a restructured convolutional neural network system using the CMP according to any one of claims 1 to 3, a) extracting a feature map by performing a convolution operation on the input image through a convolution layer; b) The feature maps extracted in step a) are reduced through the first pooling layer, and the two feature maps processed through two-stage pooling by the CMP unit and the Max Pooling unit are convolved with a size of 1 × 1 to reduce the channel. and combining them into one feature map; c) extracting a feature map by repeatedly performing a convolution operation through a plurality of convolution layers in which the feature map combined in step b) is set; d) extracting the feature maps extracted in step c) through a second pooling layer and combining the two feature maps into one feature map after extracting two feature maps performed through two-stage pooling by the CMP unit and Max Pooling; and, e) predicting and outputting the result of the final neural network through the fully combined layer using the feature map combined in step d);

Images