KR101847874B1 - Image recognition method using convolution neural network and recording medium thereof - Google Patents

Abstract

The present invention relates to an image recognition method using a convolutional neural network and a recording method thereof, which can increase speed of calculation for image classification by applying a hybrid form of a convolutional layer and a pooling layer. According to the present invention, the image recognition method is a method of recognizing an image displayed in NxN pixels using a convolutional neural network and includes a first process and a second process simultaneously performed in parallel to each other. The first process includes: step 1a of reading a first region configured of nxn (here, n<N) pixels among the NxN pixels; step 1b of performing four times of convolution operations by applying (n-1)x(n-1) kernels of ′Stride = 1′ for the first region; step 1c of extracting a first feature by consecutively performing a pooling operation with results of the convolution operations without storing the results of the convolution operations of step 1b in a shared memory; and step 1d of mapping the first feature to a convolution-pooling hybrid feature map.

Description

[0001] IMAGE RECOGNITION METHOD USING CONVOLUTION NEURAL NETWORK AND RECORDING MEDIUM THEREOF [0002]

The present invention relates to an image recognition method using a composite neural network, and more particularly, to an image recognition method using a composite neural network capable of improving a calculation speed for image classification by applying a mixture form of a composite product layer and a pooling layer And a recording medium therefor.

Neural networks based on learning data have already been used in many fields, among which Convolution Neural Network (CNN) has shown excellent performance in image recognition.

However, the articulated neural network has a problem that the resources are limited in the embedded system and the classification operation is performed with the pre-learned weight, but the operation speed is remarkably slow. To improve this, GPU (Graphics Processing Unit) can be used for GPU (general-purpose computing on Graphics Processing Units), which can be used for general-purpose calculation.

Since the CNN has many simple and repetitive operations, it can perform efficient operations by using GP-GPU that can parallelize it. Specifically, if the thread and memory approach are efficiently configured in the single instruction multiple thread (SIMT) structure of the GP-GPU, the desired result can be obtained at a higher speed than the conventional CNN method.

In order to improve the operation speed performance of CNN using the GP-GPU, it is necessary to properly allocate the threads and minimize the access time of the memory. Conventionally, an image recognition method using a composite neural network reads an entire region of an image to be recognized as one input, moves the kernel sequentially to the entire region, performs a composite product operation, Mapped to the map and then performing the pooling operation on the map is repeated a number of times to finally generate n × n (eg, 4 × 4) pooling layers. Accordingly, in the image recognition method using the conventional articulated neural network, the result of each result of the convolutional product must be stored in a shared memory, and the number of memory accesses is increased by the number of the convolutional feature maps, .

SUMMARY OF THE INVENTION The present invention has been made to solve the above problems, and it is an object of the present invention to improve a thread distribution and a memory approach in a SIMT structure of a GP-GPU, The present invention provides a method of recognizing an image using a composite neural network and a recording medium therefor.

According to an aspect of the present invention, there is provided a method for recognizing an image displayed on NxN pixels using a composite neural network, including a first process and a second process performed simultaneously and in parallel do.

The first process comprising: inputting as input a first region of n × n (where n <N) pixels of the N × N pixels; Performing four convolution operations by applying (n-1) x (n-1) kernels of Stride = 1 to the first region; Extracting a first characteristic by sequentially performing a pooling operation on the result of each of the resultant products without storing the result of each of the resultant products of the step 1b in a shared memory; And mapping the first feature to a composite product-pooling feature map.

The second process is a step of reading as input a second area of n × n pixels shifted by 'Stride = m' with respect to the first area; (B) performing four convolution operations by applying (n-1) x (n-1) kernels of Stride = 1 to the second region; Extracting a second feature by sequentially performing a pooling operation on the result of each of the resultant products, without storing the result of each of the resultant products in step 2b in a shared memory; And step 2d mapping the second feature to a composite product-pooling mix feature map.

According to the image recognition method using the composite neural network according to the present invention, each recognition target image segmented from the first input (i.e., the entire image) is divided again into a plurality of regions (i.e., first to Mth regions) , And each of the divided regions is read as an independent input to perform a combined product-pooling mixing operation of a parallel structure.

Accordingly, it is possible to omit the memory storing process of the synthesized product feature map, which is essentially required for the conventional processing method. As a result, the number of memory accesses can be reduced by the number of the conventional synthesized product feature maps, The threads can be efficiently distributed, minimizing thread computation.

As a result, the desired result can be output at a faster rate than the conventional image recognition method.

FIG. 1 is a processing flowchart showing a process of generating a composite product-pooling mixed layer according to the present invention.
FIG. 2 is a diagram illustrating a process of extracting a first feature from a first region by applying a composite product-pooling mixture operation to a first region of the present invention; FIG.
FIG. 3 is a diagram illustrating a process of dividing a recognition target image of the present invention into a plurality of images and then performing arithmetic processing in a parallel structure to generate a combined-product-pooling mixed feature map.
Fig. 4 is a diagram for explaining a nonlinear activation function used in the composite-object-based neural network of the present invention; Fig.
5 is a diagram for explaining a pulling operation of the present invention;
6 is a diagram illustrating a CNN structure of an image recognition method using a composite-object-based neural network according to the present invention.
7 is a diagram illustrating a CNN structure of an image recognition method using a conventional articulated neural network;

The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The singular expressions include plural expressions unless the context clearly dictates otherwise. In this specification, the terms "comprises" or "having" and the like refer to the presence of stated features, integers, steps, operations, elements, components, or combinations thereof, But do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, or combinations thereof.

Also, in the present specification, the term " above or above "means to be located above or below the object portion, and does not necessarily mean that the object is located on the upper side with respect to the gravitational direction. It will also be understood that when an element such as a region, plate, or the like is referred to as being "above or above another portion ", this applies not only to the presence or spacing of another portion & And the like.

Also, in this specification, when an element is referred to as being "connected" or "connected" with another element, the element may be directly connected or directly connected to the other element, It should be understood that, unless an opposite description is present, it may be connected or connected via another element in the middle.

Also, in this specification, the terms first, second, etc. may be used to describe various components, but the components should not be limited by the terms. The terms are used only for the purpose of distinguishing one component from another.

Hereinafter, preferred embodiments, advantages and features of the present invention will be described in detail with reference to the accompanying drawings.

FIG. 1 is a flowchart illustrating a process of generating a composite product-pooling mixed layer according to the present invention. 1, an image recognition method using a Convolution Neural Network (CNN) according to the present invention includes a plurality of regions S1a allocated in accordance with a predetermined interval in a recognition target image in parallel (S1b), maps the extracted features (S1c) to a composite product-pooling mixed feature map (S1d), and uses a plurality of kernels in the process to generate a plurality of composite product- And generates a composite product-pooling mixed layer made up of maps (S1e). For the sake of reference, the above-mentioned "processing of a plurality of areas" means processing by using a combination of a composite product operation and a pooling operation. Hereinafter, this is referred to as a &quot; composite product-pooling operation.

Then, a process of concurrently processing a plurality of regions assigned in accordance with a predetermined interval (Stride) concurrently with the thus-generated resultant product-pooling mixed feature map to generate a higher-level resultant product-pooling mixed feature map And is configured to output a fully connected layer in the last layer by repeating a plurality of times.

As described above, according to the present invention, it is possible to omit the memory storing process of the convolution feature map, which is essential for the conventional processing method, It is possible to reduce the number of memory accesses by the number of characteristic maps and efficiently distribute the threads, and as a result, the operation speed for image classification can be greatly improved.

FIG. 2 is a diagram illustrating a process of extracting a first feature from a first region by applying a composite product-and-pooling operation to a first region of the present invention. FIG. Parallel structure to generate a synthesized product-pooling mixed feature map. As shown in FIG.

2 and 3, one resultant product-and-pooling mixing feature map 40 (hereinafter referred to as a 'first composite product-pooling mixing feature map') includes a first process that is performed concurrently and in parallel, The second process, the third process, ..., and the M process. The first processing, the second processing, the third processing, ..., and the M processing follow the same processing method, but the input areas for the processing are different.

The example of FIG. 2 is a diagram for explaining a processing operation (for example, "first processing") of any one of the first processing, the second processing, the third processing, and the M processing, And a process of extracting a feature to be mapped to the composite product-pooling mixed feature map 40 through a composite product-pooling mixture operation on the recognition target image.

Specifically, the first process is a process for generating a first synthesized product-pooling mixed feature map 40, which includes the steps of inputting a region in detail (step 1a), performing a composite product operation (step 1b) 1c), and a feature map mapping step (step 1d).

Step 1a is a step of reading, as an input, a first region 21 having n × n (where n <N) pixel sizes among the image to be recognized 10 (Input) of N × N pixel size. In the example of FIG. 2, 'N × N' of the recognition target image 10 corresponds to '10 × 10' pixels, 'n × n' of the first region 21 corresponds to the recognition target image 10, It corresponds to a '6 × 6' pixel in the upper left of the middle.

Step 1b is a step of performing a convolution operation on the first area 21 by applying (n-1) x (n-1) kernels 30 of 'Stride = 1'. In the example of FIG. 2, '(n-1) × (n-1)' of the kernel 30 (filter) is' ((n = 6) -1) '5 × 5', and the '5 × 5' kernel 30 is subjected to a total of four convolution operations by applying "Stride = 1" to the first area 21.

If the input is f (t) in the signal processing and the kernel is g (t), the first convolution operation (hereinafter referred to as the first convolution operation) computes the overlap between the input and the kernel by inverting g . That is, in the 6 × 6 first region 21, the value at the position of each element is multiplied by the portion overlapping with the kernel located at the upper left of the 6 × 6 region, and the summation of products (SOP) do.

The second composite product operation (hereinafter referred to as a second composite product calculation) moves the kernel 30 used for the first composite product calculation to the left by one space to form the first region 21 and the kernel 30 Multiplied by the value at the position of each element in the overlapping part of the kernel 30, and outputs the result of summing the values.

The third synthesis product operation (hereinafter referred to as the third synthesis product operation) moves the kernel 30 used in the second synthesis product by one space downward to form the first region 21 and the kernel 30 The kernel at the lower part) is multiplied by the value at the position of each element in the overlapping part, and the result of summing the values is output.

The fourth synthesis product operation (hereinafter referred to as the fourth synthesis product operation) moves the kernel 30 used for the third synthesis operation to the right by one space so that the first area 21 and the kernel 30 The kernel at the lower part) is multiplied by the value at the position of each element in the overlapping part, and the result of summing the values is output.

Then, the output data is transformed into a nonlinear function through an activation function by performing each of the above-described composite product operations. The reason for using a nonlinear function is that when the activation function does not exist, the neural network degenerates into a shallow neural network no matter how deep. For example, assuming that there is a simple neuronal network with no bias, "H = WX, Y = WH" can be established by calculating the hidden layer as a matrix. This, in turn, is the result of multiplying the input by the weight between the input and the hidden and the weight between the hidden and the output, multiplied by a matrix. As a result, even if the neural network is deepened, it has a shallow neural network effect as shown in Fig. 4 (b). If you put a lot of parameters, the performance is not getting better but the speed is lowered. For this reason, it is preferable to use a nonlinear function for the activation function so that it does not become a linear function.

Step 1c is a step of extracting a first characteristic by performing a pooling operation successively with the result of each of the resultant products of step 1b without storing the result of each of the resultant products of step 1b in a shared memory, to be.

The size of the output values after performing the product multiply operation of step 1b and passing through the activation function does not become much smaller in the size of the input. In order to reduce this, pooling or subsampling operation is performed.

The pooling operation is a process of selecting a value according to a specific criterion with four values as shown in FIG. In other words, it is a process of extracting a higher level of abstracted information through a product multiply operation and compressing the size of the abstracted information so as to leave only the most important information.

Here, the compressed size is preferably 2 x 2 size compressed to 1/4 size as shown in FIG. 5, but it is not necessarily limited thereto. As for the stride, it is preferable to apply 'Stride = 2' as shown in FIG. 5, but it is not necessarily limited thereto. It is preferable to apply the Max Pooling method of outputting the maximum value, but the present invention is not limited thereto. For example, it is needless to say that an average pooling method for outputting an average value may be applied to a pooling operation.

Since the pooling operation is not multiplied with the kernel value, there is no parameter to learn, and since the number of input is outputted as it is, there is no change of channel, and it is possible to reduce the amount of variables to be calculated through such a pooling operation.

Referring to Fig. 2, step 1b has an output value by the first convolution operation, an output value by the second convolution operation, an output value by the third convolution operation, and an output value by the fourth convolution operation, (Hereinafter, referred to as 'first feature') is extracted. In the example of FIG. 2, the output value (i.e., '8' in the purple region) by the fourth composite product calculation corresponds to the first characteristic. The extracted first feature is stored in the shared memory and mapped to a feature map (step 1d) to form the first combined product-pooling mixed feature map 40 of the present invention.

The second process of the present invention is a process for generating a first synthesized product-and-pooling mixed feature map 40, which comprises the steps of inputting a region in detail (step 2a), performing a synthesis product operation (step 2b) Step 2c), and a feature map mapping step (step 2d), and is performed in a parallel structure independent of the first processing described above.

Step 2a is a step of reading as input a second region 22 of n × n pixels shifted by 'Stride = m' with respect to the first region 21 of the recognition target image 10 used in the first process to be. In the example of FIGS. 2 and 3, 'N × N' of the recognition target image 10 corresponds to '10 × 10' pixels, 'n × n' of the first region 21 corresponds to It corresponds to '6 × 6' pixel in upper left corner, and 'Stride = 2' is applied to stride. Accordingly, 'nxn' in the second area 22 corresponds to a region shifted two spaces to the right in the first area 21, that is, a '6x6' pixel in the center of the recognition target image 10 .

Step 2b is a process for convoluting the second area 22 by applying the same kernel as the first process, that is, (n-1) x (n-1) kernels 30 of 'Stride = 1' And performing an operation. In the example of FIG. 2, a '5 × 5' kernel 30 of 'Stride = 1' is applied to the second area 22 to perform a total of four convolution operations. The fourth composite product operation method of the second process is the same as that of the first process, and a detailed description thereof will be omitted.

Step 2c does not store the result of each of the products in step 2b in a shared memory but performs a pooling operation successively with the result of each of the products in step 2b, (Hereinafter referred to as &quot; second feature &quot;). The pulling operation in the step 2c is the same as the pulling operation in the step 1c of the first processing, and a detailed description thereof will be omitted.

In the case of FIG. 3, the red box area in FIG. 3 (b) corresponds to the second area 22, and the second characteristic described above with respect to the second area 22 9 'in the red color area) can be extracted. The second feature extracted in this way is mapped to the feature map (step 2d) to constitute the first synthesized product-pooling mixed feature map 40 as shown in FIG. 3 (f).

The third process of the present invention is a process for generating a first synthesized product-pooling mixed feature map 40, which includes an area input step, a composite product calculation step, a feature extraction step, and a feature map mapping step, The concrete method of each step is the same as each step of the first processing. However, the third process is shifted by 'Stride = m' with respect to the second region 22 of the second process and the third region 23 of n × n pixels is read as an input, The difference is that the features are extracted through the steps. In the example of FIG. 3, 'Stride = 2' is applied to the stride. Accordingly, 'nxn' in the third region 23 corresponds to a region shifted two spaces to the right in the second region 22, that is, a '6x6' pixel in the upper right portion of the recognition target image 10 .

(N-1) × (n-1) kernels 30 (for example, the '5 × 5' kernel in FIG. 2) of 'Stride = 1' are applied to the third area 23, (Hereinafter, referred to as 'third feature') different from the second feature by extracting the third feature from the result of the convolution operation. The extracted third feature ('5' in the green region of FIG. 3) is mapped to the feature map to form the compound-pooling mixed feature map 40 as shown in FIG. 3 (f).

The fourth process of the present invention includes an area input step, a concurrent product calculation step, a feature extraction step, and a feature map mapping step as in the third processing, and the specific method of each step is the same as each step of the first processing , But the fourth area 24 of n × n pixels is shifted by "stride = m" with respect to the third area 23 of the third process and input to the third area 23 The difference is that the synthesis product operation and feature extraction are performed using the same kernel 30 as the first process. The fourth feature ('6' in the purple region of FIG. 3) extracted through the fourth process is mapped to the feature map to construct a first synthesized product-pooling mixed feature map 40 as shown in FIG. 3 (f) .

The Mth process of the present invention is a process of processing the last area (i.e., the Mth area to be described later) for generating the first resultant multiply-and-pool mixing feature map 40, The M-th region is shifted by 'Stride = m' with respect to the M-1 region and is read as an input, and then the same kernel 30 as the first process is used for the M-th region And performs the product multiply operation and feature extraction. Here, the 'M-1 region' refers to an area allocated to a process immediately before the Mth processing. For example, if the Mth processing is the ninth processing (M = 9), the M-1 area refers to the area allocated to the eighth processing, that is, the eighth area, and the Mth area corresponds to the ninth area.

That is, the Mth region corresponds to the last region in the plurality of processing regions divided according to 'Stride = m' in the recognition target image 10. [ For example, when the recognition target image 10 of 10 × 10 size as shown in FIG. 3 and 'Stride = 2' are used as a reference, the region may be the right lower end region among the nine regions of the 10 × 10 recognition target image 10. In the example of FIG. 3, the Mth processing corresponds to the ninth processing.

In the image recognition method using the conventional articulated neural network, the entire area of the image to be recognized 10 is read as one input, and the kernel 30 is sequentially moved at a predetermined interval (Stride = n) Perform a composite product operation and map and store the output value in a composite product feature map. As a result, each result of the convolution operation must be stored in a shared memory, and the number of memory accesses is increased by the number of the convolution feature maps.

However, according to the image recognition method using the composite-object neural network of the present invention as described above, unlike the conventional method of reading the entire area of the recognition target image 10 as one input, Division into a plurality of regions (i.e., the first to M-th regions described above), and then reads the divided regions as independent inputs to perform a combined product-pooling mixing operation of a parallel structure.

(N-1) x (n-1) kernels 30 of "Stride = 1" for a plurality of areas divided from the recognition object image 10, ) Are applied in parallel to perform a combined product-pooling mixing operation, thereby mapping a feature extracted at about the same time in each region to a feature map, thereby generating a combined product-pooling mixed feature map 40 .

Accordingly, it is possible to omit the memory storing process of the convolution feature map, which is essentially required for the conventional processing method, thereby reducing the number of memory accesses by the number of the conventional result multiply feature maps, It is possible to improve the computation speed for the above.

Meanwhile, the first to Mth processes, which are image recognition methods using the composite-object-based neural network of the present invention, can be processed in parallel using GP-GPU (General-Purpose Computing on Graphics Processing Units).

In the above case, in the single instruction multiple thread (SIMT) structure of the GP-GPU, the first process is assigned to one of the plurality of threads (hereinafter referred to as a first thread) (Hereinafter, referred to as &quot; second thread &quot;) that is distinguished from the thread, and the Mth processing may be configured to be allocated to another thread that is distinguished from the preceding threads.

Accordingly, it is possible to implement a very efficient thread distribution in addition to the memory access shortening effect (that is, minimizing the memory access time) described above, so that the desired result can be outputted at a faster speed than the conventional image recognition method.

In other words, by dividing the image to be recognized 10 into a plurality of regions, and applying a combination of a composite product layer and a pooling layer to each of the plurality of regions, the amount of thread operation can be minimized.

FIG. 6 is a diagram illustrating a CNN structure of an image recognition method using a composite-object-based neural network according to the present invention. FIG. 6 illustrates a process of outputting a fore-coupling layer based on the composite product-pooling mixed layer of the present invention.

6, an image recognition method using a composite-object-based neural network according to the present invention uses 8 kernels for a recognition object image 10 of 28 × 28 size, By performing the parallel structure calculation according to the first to Mth processing methods, a first composite product-summing mixed layer composed of 12 × 12 first-order product-pooling blending feature maps 40 is generated.

Then, for each of the resultant product-pooling mixed feature maps of the first resultant product-pooling mixed layer thus generated, a parallel structure operation according to the first to Mth processing methods is performed again to extract new features And maps it to a feature map to produce a second composite product-pooling mixed layer of 4 × 4 second composite product-pooling mix feature maps.

According to the pixel size of the recognition target image 10, if necessary, the same processing method as the first to Mth processing described above is repeated a plurality of times to finally obtain the 4 × 4 size composite product-pooling blending feature maps Lt; RTI ID = 0.0 &gt; product-pooling &lt; / RTI &gt; For reference, in the example of FIG. 6, the second composite product-pooling mixed layer (i.e., the second composite product-pooling mixed layer) corresponds to a 4 × 4 composite product-pooling mixed layer.

When a 4 × 4 composite product-pooling mixed layer is generated, a fully connected layer that is fully connected to all the neurons of the composite product-pooling mixed layer is output, and the output value of the full- It is determined whether it is applicable. In the case of FIG. 6, finally, the image is classified by learning with a neural network having an input of 16 × 4 × 4 (= 256).

FIG. 7 is a diagram illustrating a CNN structure of an image recognition method using a conventional artificial neural network. FIG. 7 shows a process of outputting a foreground layer based on a conventional composite product layer and a pulling layer.

Referring to FIG. 7, in the conventional image recognition method using a composite neural network, the entire area of the image to be recognized 10 is read as one input, the kernel 30 is sequentially moved over the entire area, Operation is performed, the output value is mapped to the resultant product feature map, and then the pooling operation is performed for the resultant number of times, and finally a 4 × 4 pooling layer is generated.

As a result, in the image recognition method using the conventional artificial neural network, the result of each of the resultant products must be stored in a shared memory, and eventually the number of memory accesses increases by the number of the resultant product feature maps, There was a limit.

The following Experimental Example 1 tests the operation speed between conventional CNN and CNN based on the inventive composite product-pooling mixed layer.

Experimental Example  One

In Experimental Example 1, the feed forward operation speed was verified by using one of the MNIST [5] input 28 × 28 image and the pre-learned weight with only the feature extraction part among the CNN structure for handwriting recognition. For the experiment, I used GP-GPU with 16 warps and 16 threads in Xilinx VC707 FPGA with 256 threads in total. For reference, the feature extraction section of Experimental Example 1 refers to a composite product layer and a pooling layer in the case of conventional CNN, and refers to a composite product-pooling mixed layer in the case of CNN of the present invention.

The cumulative computation time (㎲) of the CNN computation, which is the test result according to Experimental Example 1, is summarized in Table 1 below.

Conventionally,

Conventional CNN
The feature extraction unit
The CNN of the present invention
Composite Product Layer 1

1,180

Composite-Pulling
Mixed Layer 1

1,426
Pooling Layer 1

2,180
Composite Product Layer 2

5,362

Composite-Pulling
Mixed Layer 2

4,686
Pooling Layer 2

5,610

As can be seen from Table 1, the application of the composite product-pooling layer according to the present invention reduces the memory access count by the number of feature maps of the conventional composite product layer and efficiently distributes the threads. As a result, The performance can be improved. Specifically, according to the results of Table 1 of Experimental Example 1, it can be confirmed that the cumulative cumulative time of the resultant articulated neural network according to the present invention is 924 μs less than that of the conventional art, and the time is reduced by about 16.47%.

On the other hand, the image recognition method using the composite neural network of the present invention described above and shown in the above is a method of recording a program for executing each of the above-described processes (first to Mth processes, etc.) And may be provided in the form of a recording medium readable by an electric / electronic device.

While the preferred embodiments of the present invention have been described and illustrated above using specific terms, such terms are used only for the purpose of clarifying the invention, and it is to be understood that the embodiment It will be obvious that various changes and modifications can be made without departing from the spirit and scope of the invention. Such modified embodiments should not be understood individually from the spirit and scope of the present invention, but should be regarded as being within the scope of the claims of the present invention.

10: Image to be recognized
21: first region
22: second region
23: third region
24: fourth region
25: fifth region
30: The kernel
40: Composite-Pulling Mix Feature Map

Claims (11)

A method for recognizing an image displayed on N x N pixels using a composite neural network,
A first process that is performed concurrently and in parallel; And a second process,
The first process may include:
A first region of n × n pixels (where n <N) of the N × N pixels is read as an input;
Performing four convolution operations by applying (n-1) x (n-1) kernels of Stride = 1 to the first region;
Extracting a first characteristic by sequentially performing a pooling operation on the result of each of the resultant products without storing the result of each of the resultant products of the step 1b in a shared memory; And
And mapping the first characteristic to a composite product-pooling mixed feature map,
The second process may include:
Reading a second region shifted by 'Stride = m' with respect to the first region and composed of n × n pixels as an input;
(B) performing four convolution operations by applying (n-1) x (n-1) kernels of Stride = 1 to the second region;
Extracting a second characteristic by performing a pooling operation successively with the result of each of the resultant products of step 2b without storing the result of each of the resultant products of step 2b in a shared memory; And
And mapping the second feature to a composite product-pooling feature map. 2. The method of claim 1, further comprising:
The method according to claim 1,
Further comprising a third process, a fourth process, ..., and an M process which are performed simultaneously and in parallel,
The third process is shifted by 'Stride = m' relative to the second region, and the third region consisting of n × n pixels is read as an input, and the fourth process is shifted by 'Stride = m' The Mth region is shifted by 'Stride = m' relative to the (M-1) th region and is read as an input. The Mth region is composed of n × n pixels.
(N-1) × (n-1) kernels of 'Stride = 1' are applied to each of the third region, the fourth region, the M region, The third, fourth,..., And M-th features are extracted by performing the pooling operation successively with each result of the convolution operation of each region without storing the result of each convolution operation of the region in the shared memory ; And mapping the third feature, the fourth feature, ..., and the M feature to the combined product-pooling mixed feature map, respectively.
3. The method of claim 2,
The K-th new feature is extracted by applying the same processing method as the first through M-th processes to the resultant product-pooling mixed feature map composed of the first through M-th features, and then another composite product- And mapping each of the images to the map.
The method of claim 3,
The method of claim 1, further comprising generating a composite product-pooling mixed layer comprising the composite product-pooling blend feature maps.
5. The method of claim 4,
Further comprising the step of outputting a fully connected layer that is fully connected to all the neurons of the composite product-pooling mixed layer.
6. The method of claim 5,
Further comprising the step of determining which classification corresponds to the output value of the total binding layer.
3. The method of claim 2,
Wherein the first process is performed by assigning to a first thread of the process unit,
Wherein the second process is performed by assigning to a second thread of the process unit,
Wherein the third process is performed by assigning to a third thread of the process unit,
Wherein the fourth process is performed by assigning to a fourth thread of the process unit,
And the Mth processing is performed by allocating to the Mth thread of the process unit.
8. The method of claim 7,
Wherein the process unit is GP-GPU (General-Purpose Computing on Graphics Processing Units).
The method according to claim 1,
Wherein the composite neural network uses a non-linear function as an activation function.
The method according to claim 1,
Wherein the pooling is Max Pooling or Average Pooling. &Lt; RTI ID = 0.0 &gt; 11. &lt; / RTI &gt;
In electrical and electronic devices,
10. A recording medium readable by an electric / electronic apparatus having recorded thereon a program for executing any one of the methods selected from among claims 1 to 10.

Images