Image processing method and image processing system of convolutional neural network

Abstract

The present application relates to a method and system for image processing using convolutional neural network. A system is adapted to perform a method for image processing, the method comprising obtaining input image data, first training result, second training result and interpolation correspondence table; dividing the input image data into a plurality of feature map blocks according to the total number of region interpolations; establishing a position mapping relationship to record the feature map block corresponding to the position of each region interpolation; setting the corresponding region interpolation of the feature map block according to the position mapping relationship; obtaining interpolation parameters of each feature map block according to the first training result, the second training result and the region interpolation; performing block convolution processing on each interpolation parameter and the corresponding feature map block to obtain output feature result; and merging the output feature result to obtain output image according to the position mapping relationship.

Description

TECHNICAL FIELD

[0001] A method and system for processing digital images, in particular, a method and system for processing images using a convolutional neural network. BACKGROUND

[0002] With the rise of deep neural networks, various learning algorithms have been developed in the industry. Learning algorithms have evolved from neural networks in the past to convolutional neural networks (CNNs) today. Generally speaking, learning algorithms are based on feature recognition, classification, and feedback processing. Therefore, they are often applied to feature classification and judgment of digital information. For example, image processing of digital images. SUMMARY

[0003] In view of the above, according to some embodiments, a method for processing images using a convolutional neural network includes obtaining input image data, a first training result, a second training result, and an interpolation correspondence table, the interpolation correspondence table including a plurality of region interpolations; dividing the input image data into a plurality of feature map blocks according to a total number of the region interpolations; establishing a position mapping relationship to record a position of each region interpolation corresponding to a feature map block; setting a corresponding region interpolation of the feature map block according to the position mapping relationship; obtaining an interpolation parameter for each feature map block according to the first training result, the second training result, and the region interpolation; performing block convolution processing on each interpolation parameter and the corresponding feature map block to obtain an output feature result; and merging the output feature result to obtain an output image according to the position mapping relationship. The method for processing images using a convolutional neural network divides the input image data into a plurality of feature map blocks to reduce the computational burden during image processing.

[0004] In some embodiments, the step of obtaining the interpolation correspondence table includes selecting a first convolution operation layer of the convolutional neural network; and inputting the input image data to the first convolution operation layer.

[0005] In some embodiments, the step of obtaining the output image includes selecting a second convolution operation layer of the convolutional neural network; setting the output image obtained by the first convolution operation layer as new input image data of the second convolution operation layer; and obtaining an output image of the second convolution operation layer.

[0006] In some embodiments, the step of obtaining the interpolation correspondence table includes that the first convolution operation layer and the second convolution operation layer each have a corresponding interpolation correspondence table; and selecting the corresponding interpolation correspondence table according to the first convolution operation layer or the second convolution operation layer.

[0007] In some embodiments, the step of dividing the input image data into a plurality of feature map blocks according to the total number of region interpolations comprises calculating a first area of each feature map block according to the number of interpolations and the input image data; and expanding the first area of each feature map block to a second area such that there is an overlapping region between two adjacent feature map blocks.

[0008] In some embodiments, the step of merging the interpolation parameters according to the position mapping relationship to obtain the output image comprises superimposing overlapping regions corresponding to two adjacent interpolation parameters according to the position mapping relationship; and obtaining the output image.

[0009] In some embodiments, the step of obtaining the interpolation parameter of each feature map block according to the first training result, the second training result, and the region interpolation comprises setting a corresponding first temporary region interpolation of each feature map block according to the first training result and the position mapping relationship; setting a corresponding second temporary region interpolation of each feature map block according to the second training result and the position mapping relationship; and obtaining the interpolation parameter of a corresponding position according to the first training result, the second training result, the first temporary region interpolation, and the second temporary region interpolation.

[0010] In some embodiments, the image processing system comprises a storage unit and a processor. The storage unit stores a convolutional neural network program, input image data, at least one interpolation corresponding table, a first training result, a second training result, and an output image; the interpolation corresponding table comprises a plurality of region interpolations, and the total number of the region interpolations is a number of interpolations; the processor executes the convolutional neural network program according to the input image data and the interpolation corresponding table; the processor divides the input image data into a plurality of feature map blocks according to the number of interpolations and a position mapping relationship; each feature map block sets a corresponding region interpolation according to the position mapping relationship; the processor calculates an interpolation parameter of each feature map block according to the first training result, the second training result, and the region interpolation; and the processor merges the interpolation parameters according to the position mapping relationship to obtain the output image. The convolutional neural network image processing system divides the input image data into a plurality of feature map blocks, so as to reduce the computational burden of the processor when processing the image.

[0011] In some embodiments, the convolutional neural network program comprises a first convolutional operation layer and a second convolutional operation layer, the output of the first convolutional operation layer is connected to the second convolutional operation layer, and the first convolutional operation layer and the second convolutional operation layer each have an interpolation corresponding table.

[0012] In some embodiments, the processor calculates a first area of each feature map block according to the number of interpolations and the input image data, expands the first area of each feature map block to a second area such that there is an overlapping region between two adjacent feature map blocks, calculates an interpolation parameter of the overlapping region according to the first training result, the second training result, and the region interpolation, and obtains the output image according to the position mapping relationship.

[0013] In summary, the image processing method and system based on convolutional neural networks proposes using multiple sets of training results and region interpolation to adjust new interpolation parameters, and further obtaining the output features of feature maps based on the new interpolation parameters. Furthermore, by segmenting the input image data, the processor can concentrate its computational resources on processing feature maps, thereby reducing the computational burden on large input samples. Attached Figure Description

[0014] [ Figure 1A [This is a functional block diagram of an embodiment of an image processing system.]

[0015] [ Figure 1B This is a schematic diagram of an embodiment of a single set of convolutional operation layers in an image processing system.

[0016] [ Figure 1C This is a schematic diagram of an embodiment of interpolation correspondence table and region interpolation in an image processing system.

[0017] [ Figure 2 [Image] is a flowchart illustrating an embodiment of an image processing method.

[0018] [ Figure 3A [This is a schematic diagram of an embodiment of the input image data and feature block segmentation of an image processing method.]

[0019] [ Figure 3B [This is a schematic diagram illustrating the generation of interpolation parameters in an embodiment of an image processing method.]

[0020] [ Figure 3C [This is a schematic diagram illustrating the generation of output feature results in an embodiment of an image processing method.]

[0021] [ Figure 4A [This is a schematic diagram of multiple sets of convolutional operation layers in an embodiment of an image processing method.]

[0022] [ Figure 4B This is a schematic diagram illustrating the operation of each convolutional operation layer in an embodiment of an image processing method.

[0023] [ Figure 4C This is a schematic diagram illustrating the operation of each convolutional operation layer in an embodiment of an image processing method.

[0024] [ Figure 5 [This is an extended schematic diagram of a feature block in an embodiment of an image processing method.]

[0025] [ Figure 6 [This is a schematic diagram of multiple training results and corresponding region interpolation in an embodiment of an image processing method.] Detailed Implementation

[0026] Please refer to Figure 1A The diagram shown is a functional block diagram of an embodiment of an image processing system. The image processing system 100 can be implemented using a personal computer or server, or it can be used in an embedded controller or a system-on-a-chip. The image processing system 100 includes a storage unit 110 and a processor 120. The storage unit 110 stores a convolutional neural network program (hereinafter referred to as CNN program 111), input image data 112, at least one interpolation mapping table 113, a first training result 114, a second training result 115, and an output image 116. In one embodiment, the internal processing of one of the convolutional operation layers is described as follows: Figure 1B As shown in the dashed box.

[0027] The CNN program 111 sets up an interpolation correspondence table 113 based on the computational processing layer. The interpolation correspondence table 113 has multiple region interpolations 117; please refer to [reference needed]. Figure 1C As shown. For ease of explanation, Figure 1C The system uses a two-dimensional array as the set and representation of region interpolation 117. The top-left corner of the interpolation correspondence table 113 is the origin, and α(x, y) represents the region interpolation 117. x and y correspond to the numbers on the horizontal and vertical axes of the interpolation correspondence table 113, respectively. For example, the region interpolation α(1, 0) corresponds to the second row and first column of the interpolation correspondence table 113. Furthermore, the interpolation correspondence table 113 can also be represented as a one-dimensional or three-dimensional array.

[0028] Input image data 112 is the image file that the CNN program 111 intends to process. Output image 116 is the output result after processing by the CNN program 111. The output result is determined by the convolutional operation layers.

[0029] The first training result 114 is a training sample generated by the first digital image processing program. The second training result 115 is a training sample generated by the second digital image processing program. The first and second digital image processing programs can be, but are not limited to, enhanced super-resolution, denoise, image-to-image translation, style transfer, or semantic translation, and the first and second digital image processing programs are different image processing programs, or similar image processing programs generated using different training data.

[0030] Referring to Figure 2 FIG. 1 shows a flowchart of an embodiment of an image processing method using a convolutional neural network. The image processing method using a convolutional neural network includes the following steps:

[0031] Step S210: Obtain input image data, a first training result, a second training result, and an interpolation correspondence table, wherein the interpolation correspondence table includes a plurality of region interpolations, and the total number of the region interpolations is an interpolation number.

[0032] Step S220: Divide the input image data into a plurality of feature map blocks according to the interpolation number.

[0033] Step S230: Establish a position mapping relationship to record the position of each region interpolation corresponding to a feature map block.

[0034] Step S240: Set the corresponding region interpolation of the feature map block according to the position mapping relationship.

[0035] Step S250: Calculate an interpolation parameter of each feature map block according to the first training result, the second training result, and the region interpolation.

[0036] Step S260: Perform block convolution processing on each interpolation parameter and the corresponding feature map block to obtain an output feature result.

[0037] Step S270: Merge the output feature result according to the position mapping relationship to obtain an output image.

[0038] First, the processor 120 executes the CNN program 111 and loads the input image data 112, the first training result 114, the second training result 115, and the interpolation correspondence table 113 from the storage unit 110. The interpolation correspondence table 113 has a plurality of region interpolations 117, as shown in FIG. 1. Figure 1C The total number of the region interpolations 117 is an interpolation number. The processor 120 divides the input image data 112 into an equal number of feature map blocks 118 according to the interpolation number, as shown in FIG. 2. Figure 3A The size of the feature map block 118 is determined according to the input image data 112 and the interpolation number. In Figure 3A , the input image data 112 and the interpolation correspondence table 113 are shown in a two-dimensional array, and the size of the feature map block 118 is n*m pixels, where n and m are positive integers. Figure 3A The upper right part is the input image data 112 that is not divided, Figure 3A The upper left part is the interpolation correspondence table 113, Figure 3A The lower part is the input image data 112 after division (i.e., the set of feature map blocks 118). Figure 3AThe arrows represent the corresponding positions of the input image data 112 and the interpolation correspondence table 113.

[0039] During the segmentation of feature map blocks 118, the CNN program 111 establishes a positional mapping relationship between the positions of each region interpolation 117 in the interpolation correspondence table 113 and the corresponding region interpolation 117. Since the input image data 112 is segmented according to the interpolation correspondence table 113, the arrangement of feature map blocks 118 and the arrangement of region interpolation 117 in the interpolation correspondence table 113 can be considered the same. Therefore, each region interpolation 117 can be mapped to a corresponding feature map block 118. The CNN program 111 assigns the corresponding region interpolation 117 to the feature map block 118 according to the positional mapping relationship. In other words, the positional mapping relationship is the location of the feature map block 118 and the corresponding region interpolation 117.

[0040] Next, CNN program 111 calculates the interpolation parameters 210 for each feature map block 118 based on the first training result 114, the second training result 115, and the region interpolation 117. Here, it is assumed that the content of the first training result 114 is "A", hereinafter referred to as the first training result A, A = {a0, a1, ..., an}; and the content of the second training result 115 is "B", hereinafter referred to as the second training result B, B = {b0, b1, ..., bn}. The region interpolation 117 is α(u, v), u ∈ {0~x}, v ∈ {0~y}. Feature map block T(u, v) 118 obtains the corresponding region interpolation 117 based on the u and v positions. CNN program 111 performs interpolation calculations based on the first training result A, the second training result B, and the region interpolation α(u, v). Please cooperate... Figure 3B As shown. CNN program 111 obtains the corresponding region interpolation α(u,v) based on feature map T(u,v) 118. CNN program 111 performs interpolation calculations on the first training result A and the second training result B with the region interpolation α(u,v) to obtain the interpolation parameters I. n (u, v), as shown in Equation 1 below:

[0041] Interpolation parameter I n (u, v) = A*α (u, v) + B* (1-α (u, v)) Formula 1 In other words, the interpolation parameter I is composed of array elements of {i0, i1, ..., in}. Equation 1 represents the set of interpolation parameters I, and the interpolation parameters for this set can be expressed as:

[0042] I n (u, v) = a n *α(u, v)+b n *(1-α(u, v)) Equation 2

[0043] In the above formula 1, 2, the first training result A and the second training result B respectively include a plurality of sets of interpolation parameters 210. Taking the first training sample A as an example, the interpolation parameters 210 of each set are as follows:

[0044] I0(u, v) = a0* a(u, v) + b0* (1 - a(u, v));

[0045] I1(u, v) = a1* a(u, v) + b1* (1 - a(u, v)); ...

[0047] In(u, v) = an* a(u, v) + bn* (1 - a(u, v)).

[0048] The CNN program 111 repeatedly obtains the corresponding region interpolation 117 for other feature map blocks 118 and calculates the corresponding interpolation parameters 210. After the CNN program 111 obtains the interpolation parameters I(u, v) of the feature map block T(u, v) 118, the CNN program 111 performs block convolution processing 310 on the feature map block 118 according to the interpolation parameters I(u, v) and obtains an output feature result 320. The CNN program 111 repeatedly calculates the interpolation parameters 210 for other feature map blocks 118 and obtains the output feature results 320 of the feature map blocks 118.

[0049] Finally, the CNN program 111 merges all the output feature results 320 according to the position mapping relationship, as shown in Figure 3C . In other words, the CNN program 111 writes the corresponding output feature results 320 into the corresponding positions of the output image 116 according to the feature map blocks 118 and the position mapping relationship. In Figure 3C , the feature map blocks 118 in the input image data 112, the interpolation parameters 210 of the corresponding positions, and the output feature results 320 of the corresponding positions are represented by gray blocks.

[0050] In some embodiments, the CNN program 111 has multiple convolution operation layers, and two convolution operation layers are connected to each other, as shown in Figure 4A and Figure 4B . In order to distinguish the convolution operation layers connected in front and back, they are respectively referred to as the first convolution operation layer 410 and the second convolution operation layer 420. The first convolution operation layer 410 is used to process the first input image data 411, and the second convolution operation layer 420 is used to process the second input image data 421. The first convolution operation layer 410 is correspondingly provided with the first interpolation corresponding table 412. The second convolution operation layer 420 is correspondingly provided with the second interpolation corresponding table 422. The number of interpolations of the first interpolation corresponding table 412 does not necessarily equal the number of interpolations of the second interpolation corresponding table 422.

[0051] The first interpolation correspondence table 412 and the second interpolation correspondence table 422 are both two-dimensional array representations. The interpolation quantity of the first interpolation correspondence table 412 is Z1, where Z1 = X1*Y1, X1 is the number of elements in the horizontal axis of the first interpolation correspondence table 412, and Y1 is the number of elements in the vertical axis of the first interpolation correspondence table 412. The interpolation quantity of the second interpolation correspondence table 422 is Z2, where Z2 = X2*Y2, X2 is the number of elements in the horizontal axis of the second interpolation correspondence table 422, and Y2 is the number of elements in the vertical axis of the second interpolation correspondence table 422, and X1≠X2, Y1≠Y2. The region interpolation 117 of the first interpolation correspondence table 412 is represented by α(u, v), and the region interpolation 117 of the second interpolation correspondence table 422 is represented by β(u, v), where u∈{0~x}, v∈{0~y}.

[0052] First, the CNN program 111 loads the first input image data 411 into the first convolution operation layer 410. The CNN program 111 calls the first interpolation correspondence table 412 according to the first convolution operation layer 410. The CNN program 111 divides the first input image data 411 into a plurality of feature map blocks 118 according to the first interpolation correspondence table 412. The CNN program 111 obtains the region interpolation α(u, v) corresponding to each feature map block 118 according to the first interpolation correspondence table 412 and the position mapping relationship. The CNN program 111 performs the aforementioned calculation on each region interpolation 117, the first training result A, and the second training result B to obtain the interpolation parameter 210 corresponding to each feature map block 118.

[0053] The CNN program 111 performs block convolution processing 310 on the feature map block 118 and the corresponding interpolation parameter 210 to obtain an output feature result 320. When the block convolution processing 310 on all feature map blocks 118 and interpolation parameters 210 is completed, the CNN program 111 sequentially combines the output feature results 320 according to the position mapping relationship to obtain a first output image (corresponding to the second input image data 421 below).

[0054] Since the first convolution operation layer 410 is connected to the second convolution operation layer 420, the first interpolation correspondence table 412 and the second interpolation correspondence table 422 can be simultaneously referred to Figure 1B and Figure 4B Therefore, the CNN program 111 takes the first output image as the second input image data 421. The CNN program 111 calls the second interpolation correspondence table 422 and performs division and other processing on the second input image data 421. In the processing of the second convolution operation layer 420, the CNN program 111 can select other training results in addition to the second interpolation correspondence table 422. Figure 4B In the processing of the second convolution operation layer 420 of the second interpolation correspondence table 422, the CNN program 111 selects the first training result C and the second training result D as the calculation reference of the interpolation parameter 210, as shown in Figure 4CThe CNN program 111 obtains the second output image 430 outputted by the second convolution operation layer 420.

[0055] In other words, the CNN program 111 can perform the aforementioned processing on the plurality of convolution operation layers connected in series. The CNN program 111 takes the output image 116 of the first convolution operation layer 410 as the input image data 112 of the next convolution operation layer. After passing through the plurality of convolution operation layers, the boundary stitching of the feature map blocks 118 can be integrated so that the boundaries of two adjacent feature map blocks 118 do not cause discontinuous visual gaps.

[0056] In some embodiments, the following variation processing can be added in the process of dividing the input image data 112 into the feature map blocks 118. When the CNN program 111 divides the input image data 112 according to the interpolation number, the CNN program 111 calculates the size of the first area of each feature map block 118 according to the interpolation number and the input image data 112. Then, the CNN program 111 adjusts the area of the feature map block 118 so that the first area of the feature map block 118 increases to the second area.

[0057] For example, if the first area of the feature map block 118 is 3*5 pixels in size and the convolution network size is 3*3. The CNN program 111 will expand the first area to 5*7 pixels in size according to the convolution network size, and keep the output to be 3*5 pixels in size after passing through the CNN program 111, to ensure that the output feature map block 118 does not have the phenomenon of offset or overlap. The expansion method can be padding or taking the pixels of the adjacent blocks.

[0058] The CNN program 111 obtains the region interpolation 117 according to the feature map block 118 and the position mapping relationship, and obtains the interpolation parameter 210 of the corresponding position. The CNN program 111 expands according to the convolution scale size, and performs block convolution processing 310 on each interpolation parameter 210 and the corresponding feature map block 118 to obtain the output feature result 320.

[0059] The CNN program 111 obtains the region interpolation 117 according to the feature map block 118 and the position mapping relationship, and obtains the interpolation parameter 210 of the corresponding position. The CNN program 111 performs block convolution processing 310 on each interpolation parameter 210 and the corresponding feature map block 118 to obtain the output feature result 320. Since there are overlapping regions 510 (as shown) between the feature map blocks 118, the CNN program 111 performs corresponding processing on the overlapping regions 510 according to the interpolation parameter 210 during the merging of the output image 116, and obtains the output feature result 320. Figure 5

[0060] ​In some embodiments, the CNN program 111 can set multiple sets of region interpolation 117 and generate corresponding interpolation correspondence table 620. The interpolation correspondence table 620 has multiple sets of training samples and corresponding region interpolations 117. The CNN program 111 calculates the interpolation parameters 610 according to the respective training results. In other words, the CNN program 111 can perform interpolation calculation of multiple sets of training results simultaneously in the same layer convolution operation.

[0061] Region interpolation α in this embodiment n where n∈{A, B, C, D}. Region interpolation α n corresponds to the first training result A, the second training result B, the third training result C, and the fourth training result D. Further, the first training result A sets the first temporary region interpolation corresponding to each feature map block according to the position mapping relationship; the second training result B sets the second temporary region interpolation; the third training result C sets the third temporary region interpolation; and the fourth training result D sets the fourth temporary region interpolation, and please simultaneously cooperate Figure 6 However, the number of training samples is not limited to this, and more training samples and difference calculations can be performed in an environment where the computing power allows. Therefore, the interpolation parameter 210 is as follows:

[0062] I n (u, v) = A*α A (u, v) + B*α B (u, v) + C*α C (u, v) + D*α D (u, v) Equation 3

[0063] Figure 6 The upper block represents each set of training results (training results A, B, C, and D, respectively) in the interpolation parameter I n (u, v). Figure 6 The lower dashed box contains each region interpolation 117 of the first training result A, the second training result B, the third training result C, and the fourth training result D.

[0064] The image processing method and system of the convolutional neural network can adjust the multiple sets of training results and region interpolations 117 to new interpolation parameters 210, and further obtain the output features of the feature map block 118 according to the new interpolation parameters 210. In addition, through the segmentation processing of the input image data 112, the processor 120 can concentrate the operation resources on the processing of the feature map block 118, thereby reducing the operation burden of the processor 120 for large input samples. In addition, the processor 120 can perform the foregoing processing on each convolution layer of the convolutional neural network, so as to adjust the boundary splicing discontinuity of the feature map block.

[0065]

Symbol Description

[0066] 100: image processing system

[0067] 110: storage unit

[0068] 111: CNN program

[0069] 112: input image data

[0070] 113, 620: interpolation correspondence table

[0071] 114, A, C: first training result

[0072] 115, B, D: second training result

[0073] 116: output image

[0074] 117: region interpolation

[0075] 118: feature map block

[0076] 120: processor

[0077] 210, 610: interpolation parameter

[0078] 310: block convolution processing

[0079] 320: output feature result

[0080] 410: first convolution operation layer

[0081] 411: first input image data

[0082] 412: first interpolation correspondence table

[0083] 420: second convolution operation layer

[0084] 421: second input image data

[0085] 422: second interpolation correspondence table

[0086] 430: second output image

[0087] 510: overlapping region

[0088] S210, S220, S230, S240, S250, S260, S270: step

Claims

1. A method for image processing of a convolutional neural network, comprising: obtaining an input image data, a first training result, a second training result, and an interpolation correspondence table including a plurality of region interpolations; dividing the input image data into a plurality of feature map blocks according to a total number of the plurality of region interpolations; establishing a position mapping relationship to record a position of each of the region interpolations corresponding to the feature map blocks; setting the region interpolations corresponding to the feature map blocks according to the position mapping relationship; obtaining an interpolation parameter of each of the feature map blocks according to the first training result, the second training result, and the plurality of region interpolations; performing a block convolution process on each of the interpolation parameters and the corresponding feature map blocks to obtain an output feature result; and merging a plurality of the output feature results to obtain an output image according to the position mapping relationship. The step of obtaining the interpolation parameter of each of the feature map blocks according to the first training result, the second training result, and the plurality of region interpolations comprises: setting a first temporary region interpolation corresponding to each of the feature map blocks according to the first training result and the position mapping relationship; setting a second temporary region interpolation corresponding to each of the feature map blocks according to the second training result and the position mapping relationship; and obtaining the interpolation parameter of the corresponding position according to the first training result, the second training result, the first temporary region interpolation, and the second temporary region interpolation. 2.The method of claim 1, wherein before the step of obtaining the interpolation correspondence table, comprising: selecting a first convolution operation layer of a convolutional neural network; and inputting the input image data to the first convolution operation layer. 3.The method of claim 2, wherein the step of obtaining the output image comprises: selecting a second convolution operation layer of the convolutional neural network; setting the output image obtained by the first convolution operation layer as new input image data of the second convolution operation layer; and obtaining the output image of the second convolution operation layer. 4.The method of claim 3, wherein before the step of obtaining the interpolation correspondence table, comprising: the first convolution operation layer and the second convolution operation layer each have a corresponding interpolation correspondence table; and selecting the corresponding interpolation correspondence table according to the first convolution operation layer or the second convolution operation layer. 5.The method of claim 1, wherein the step of dividing the input image data into a plurality of feature map blocks according to the total number of the plurality of region interpolations comprises: calculating a first area of each of the feature map blocks according to the number of interpolations and the input image data; and extending the first area of each of the feature map blocks to a second area, so that there is an overlapping region between two adjacent feature map blocks. 6.The method of claim 5, wherein the step of merging a plurality of the output feature results to obtain the output image according to the position mapping relationship comprises: superimposing the overlapping region corresponding to two adjacent interpolation parameters according to the position mapping relationship; and ​ ​ ​ ​ ​ obtaining the output image.

7. An image processing system of a convolutional neural network, comprising: a storage unit storing a convolutional neural network program, an input image data, at least one interpolation correspondence table, a first training result, a second training result, and an output image; the interpolation correspondence table comprising a plurality of region interpolations, a total number of the plurality of region interpolations being an interpolation number; and a processor executing the convolutional neural network program according to the input image data and the interpolation correspondence table, the processor dividing the input image data into a plurality of feature map blocks according to the interpolation number and a position mapping relationship, each of the feature map blocks being set with a corresponding region interpolation according to the position mapping relationship, the processor calculating an interpolation parameter of each of the feature map blocks according to the first training result, the second training result, and the plurality of region interpolations, the processor merging output feature results calculated based on the interpolation parameters to obtain the output image according to the position mapping relationship, wherein the calculating of the interpolation parameter of each of the feature map blocks according to the first training result, the second training result, and the plurality of region interpolations comprises: setting a first temporary region interpolation of each of the feature map blocks according to the first training result and the position mapping relationship; setting a second temporary region interpolation of each of the feature map blocks according to the second training result and the position mapping relationship; and calculating the interpolation parameter of a corresponding position according to the first training result, the second training result, the first temporary region interpolation, and the second temporary region interpolation.

8. The image processing system of the convolutional neural network according to claim 7, wherein the convolutional neural network program comprises a first convolutional operation layer and a second convolutional operation layer, the first convolutional operation layer being output-connected to the second convolutional operation layer, the first convolutional operation layer and the second convolutional operation layer each having the interpolation correspondence table.

9. The image processing system of the convolutional neural network according to claim 7, wherein two adjacent feature map blocks have an overlapping region, the processor calculating the interpolation parameter of the overlapping region according to the first training result, the second training result, and the plurality of region interpolations, the processor obtaining the output image according to the position mapping relationship.

Images