TWI776489B - Electronic device and method for document segmentation - Google Patents

Abstract

An electronic device and a method for document segmentation are provided. The method includes: obtaining a first feature map and a second feature map corresponding to an original document; performing a first upsampling on the second feature map to generate a third feature map; concating the first feature map and the third feature map to generate a fourth feature map; inputting the fourth feature map to a first inverted residual block (IRB) and performing a first atrous convolution operation based on a first dilation rate to generate a fifth feature map; inputting the fourth feature map to a second IRB and performing a second atrous convolution operation based on a second dilation rate to generate a sixth feature map; concating the fifth feature map and the sixth feature map to generate a seventh feature map; performing a convolution operation on the seventh feature map to generate a segmented document.

Description

Electronic device and method for file segmentation

The present invention relates to an electronic device and method for file segmentation.

Currently, document segmentation is a technique that has received attention in the field of semantic segmentation. Document segmentation can be used to identify and label individual objects in a document (eg textual content, images or tables). Although many deep learning-based file segmentation methods have been proposed, the results produced by these methods are still limited by the amount of computing resources. For example, a convolutional neural network with fewer convolutional layers may not be able to label objects in a file very clearly. Accordingly, how to propose a file segmentation method that can achieve better results with less computing resources is one of the goals of those in the art.

The present invention provides an electronic device and method for file segmentation, which can utilize a small amount of computing resources to perform file segmentation on a file to generate a segmented file.

An electronic device for file division of the present invention includes a processor, a storage medium and a transceiver. The transceiver receives the original file. The storage medium stores the neural network model. The processor is coupled to the storage medium and the transceiver, and accesses and executes a neural network model, wherein the neural network model includes a first model, wherein the first model is configured to perform: obtaining a first size corresponding to the first size of the original file a feature map and a second feature map of a second size, wherein the first size is larger than the second size; performing a first upsampling on the second feature map to generate a third feature map of a third size, wherein the third size is equal to the first size; concatenate the first feature map and the third feature map to generate a fourth feature map; input the fourth feature map to the first inverse residual block to generate a first output, and perform a first dilation rate based on the first output The first atrous convolution operation of to generate the fifth feature map; the fourth feature map is input to the second inverse residual block to generate the second output, and the second dilation rate based second atrous convolution is performed on the second output product operation to generate a sixth feature map, wherein the second dilation rate is different from the first dilation rate; connecting the fifth feature map and the sixth feature map to generate a seventh feature map; and performing a first convolution operation on the seventh feature map to generate a segmented file, wherein the processor outputs the segmented file through the transceiver.

In an embodiment of the present invention, the above-mentioned neural network model further includes a second model, wherein the second model is configured to perform: performing a second upsampling on the second feature map to generate an eighth feature map of a fourth size , where the fourth dimension is equal to the first dimension; concatenate the first feature map and the eighth feature map to generate a ninth feature map; and perform a second convolution operation on the ninth feature map to generate an output feature map.

In an embodiment of the present invention, the above-mentioned first model corresponds to a first loss function, wherein the second model corresponds to a second loss function, wherein the processor adds the first loss function and the second loss function to generate a third loss function a loss function, wherein the processor trains the first model and the second model according to the third loss function.

In an embodiment of the present invention, the above-mentioned neural network model further includes a coding convolutional network, wherein the coding convolutional network includes a first coding convolutional layer and a second coding convolutional layer, wherein the coding convolutional network is configured to perform: generating a first encoded feature map according to the original file and the first encoded convolutional layer; and generating a second encoded feature map based on the first encoded feature map and the second encoded convolutional layer.

In an embodiment of the present invention, the above-mentioned neural network model further includes a decoding convolutional network, wherein the decoding convolutional network includes a first decoding layer and a second decoding layer, wherein the first decoding layer includes a second encoding volume a convolutional layer and a decoding convolutional layer corresponding to the second encoding convolutional layer, wherein the decoding convolutional network is configured to perform: generating a second feature map based on the second encoding feature map and the first decoding layer; and generating a second feature map based on the second feature map and the first decoding layer The second decoding layer produces the first feature map.

In one embodiment of the present invention, the first model described above is further configured to perform: adding the first feature map and the third feature map to generate a tenth feature map; and connecting the tenth feature map, the first feature map, and a third feature map to generate a fourth feature map.

In one embodiment of the present invention, the first model described above is further configured to perform: adding the fifth feature map and the sixth feature map to generate an eleventh feature map; and connecting the fifth and sixth feature maps and the eleventh feature map to generate the seventh feature map.

In one embodiment of the present invention, the first model described above is further configured to perform: performing a first convolution operation on the seventh feature map to generate a twelfth feature map; and inputting the twelfth feature map to extrusion and incentivizing the network to produce segmented files.

In an embodiment of the present invention, the above-mentioned first encoding convolutional layer performs a moving reverse bottleneck convolution on the original file to generate a first encoding feature map.

A method of the present invention for file segmentation, comprising: obtaining an original file and a neural network model, wherein the neural network model includes a first model, wherein the first model is configured to perform: obtaining a first size corresponding to the original file and a second feature map of a first size of a first dimension; connecting the first feature map and the third feature map to generate a fourth feature map; inputting the fourth feature map to a first inverse residual block to generate a first output, and performing a a first dilated convolution operation of the dilation rate to generate a fifth feature map; the fourth feature map is input to a second inverse residual block to generate a second output, and a second dilation rate based second output is performed on the second output atrous convolution operation to generate a sixth feature map, where the second dilation rate is different from the first dilation rate; concatenating the fifth feature map and the sixth feature map to generate a seventh feature map; and performing the first convolution on the seventh feature map product operations to produce a split file; and outputting the split file.

Based on the above, the architecture of the neural network model proposed by the present invention can produce better results than traditional file segmentation methods under the condition of using less computing resources.

In order to make the content of the present invention more comprehensible, the following specific embodiments are given as examples according to which the present invention can indeed be implemented. Additionally, where possible, elements/components/steps using the same reference numerals in the drawings and embodiments represent the same or similar parts.

FIG. 1 is a schematic diagram of an electronic device 100 for file division according to an embodiment of the present invention. The electronic device 100 may include a processor 110 , a storage medium 120 and a transceiver 130 .

The processor 110 is, for example, a central processing unit (CPU), or other programmable general-purpose or special-purpose micro control unit (micro control unit, MCU), microprocessor (microprocessor), digital signal processing digital signal processor (DSP), programmable controller, application specific integrated circuit (ASIC), graphics processor (graphics processing unit, GPU), image signal processor (image signal processor, ISP) ), image processing unit (IPU), arithmetic logic unit (ALU), complex programmable logic device (CPLD), field programmable gate array (field programmable gate array) , FPGA) or other similar elements or a combination of the above. The processor 110 may be coupled to the storage medium 120 and the transceiver 120 , and access and execute a plurality of modules and various application programs stored in the storage medium 120 .

The storage medium 120 is, for example, any type of fixed or removable random access memory (random access memory, RAM), read-only memory (ROM), and flash memory (flash memory). , hard disk drive (HDD), solid state drive (solid state drive, SSD) or similar components or a combination of the above components for storing a plurality of modules or various applications executable by the processor OOO. In this embodiment, the storage medium 120 can store the neural network model 200 for file segmentation of the original file.

The transceiver 130 transmits and receives signals in a wireless or wired manner. Transceiver 130 may also perform operations such as low noise amplification, impedance matching, frequency mixing, up or down frequency conversion, filtering, amplification, and the like. The electronic device 100 may receive the original file through the transceiver 130, so as to use the neural network model in the storage medium 120 to perform file segmentation on the original file.

FIG. 2 is a schematic diagram of a neural network model 200 according to an embodiment of the present invention. The neural network model 200 may include an encoding convolutional network 210, a decoding convolutional network 220, a first model 230, and a second model 240, wherein the first model 230 may include a densely joint pyramid module (DJPM) 231. In one embodiment, the first model 230 may further include a squeeze-and-excitation network (SENet) 232 . The neural network model 200 may receive the raw file 30 and convert the raw file 30 into a processed file. FIG. 3 illustrates a schematic diagram of an original document 30 and a processed document according to an embodiment of the present invention. The processed files may include the cut file 40 output by the first model 230 and the cut file 50 output by the second model 240 . As can be seen from FIG. 3 , the cut document 40 (or cut document 50 ) can clearly identify the different items in the original document 30 . In other words, the performance of the file segmentation of the neural network model 200 is excellent.

Referring to FIG. 2 , the coding convolutional network 210 may include a plurality of coding convolutional layers, wherein the number of the plurality of coding convolutional layers can be adjusted according to requirements, which is not limited in the present invention. In this embodiment, the encoding convolutional network 210 may include an encoding convolutional layer 211, an encoding convolutional layer 212, an encoding convolutional layer 213, an encoding convolutional layer 214, an encoding convolutional layer 215, an encoding convolutional layer 216, an encoding convolutional layer 217, and an encoding convolutional layer 217. Convolutional layer 218.

The encoding convolution layer 211 may receive the original document 30 and perform a convolution operation on the original document 30 to generate an encoded feature map. The encoding convolutional layer 212 may receive the encoded feature map output by the encoding convolutional layer 211, and perform a convolution operation on the encoded feature map output by the encoding convolutional layer 211 to generate a new encoded feature map. In a similar manner, the coding convolutional layer in the coding convolutional network 210 may receive the coding feature map output by the previous coding convolutional layer and generate a new coding feature map according to the received coding feature map. After the convolution operation of multiple encoding convolutional layers, the encoding convolutional layer 218 may perform a convolution operation on the encoded feature map output by the encoding convolutional layer 217 to generate a new encoded feature map.

The multiple encoded convolutional layers in the encoded convolutional network 210 may correspond to different sizes. In other words, the encoded feature maps output by different encoding convolutional layers may have different sizes. For example, the size of the encoded feature map output by the encoding convolutional layer 211 may be different from the size of the encoded feature map output by the encoding convolutional layer 212 . The coding convolutional network 210 can utilize multiple coding convolutional layers of different sizes to extract important features of the original document 30 in time or space at multiple scales.

In one embodiment, the plurality of encoded convolutional layers in the encoded convolutional network 210 may be mobile inverted bottleneck convolution (MBConv) layers. Taking the encoding convolutional layer 211 as an example, the encoding convolutional layer 211 may perform a moving reverse bottleneck convolution on the original document 30 to generate an encoded feature map. Taking the encoded convolutional layer 212 as an example, the encoded convolutional layer 212 may perform a moving reverse bottleneck convolution on the encoded feature map output by the encoded convolutional layer 211 to generate a new encoded feature map.

The decoding convolutional network 220 can include a plurality of decoding layers, wherein the number of the plurality of decoding layers can be adjusted according to requirements, which is not limited in the present invention. In this embodiment, the number of multiple decoding layers may be the number of multiple encoding convolutional layers in the encoding convolutional network 210 minus 1. The decoding convolutional network 220 may include a decoding layer 221 , a decoding layer 222 , a decoding layer 223 , a decoding layer 224 , a decoding layer 225 , a decoding layer 226 , and a decoding layer 227 .

One or more decoding layers in decoding convolutional network 220 may correspond to one or more encoding convolutional layers in encoding convolutional network 210 . In this embodiment, the decoding layer 221 may correspond to the encoding convolutional layer 217 . Decoding layer 222 may correspond to encoding convolutional layer 216 . Decoding layer 223 may correspond to encoding convolutional layer 215 . Decoding layer 224 may correspond to encoding convolutional layer 214 . Decoding layer 225 may correspond to encoding convolutional layer 213 . Decoding layer 226 may correspond to encoding convolutional layer 212 . Decoding layer 227 may correspond to encoding convolutional layer 211 .

In the decoding convolutional network 220, one or more decoding layers that are closer to the encoding convolutional network 210 (ie, one or more decoding layers that are closer to the input of the encoding convolutional network 210) may be Contains encoded convolutional layers. The coded convolutional layers in the decoding layers can be at the input or output of the decoding layers. The decoding layer may be a concatenation of an encoding convolutional layer and a decoding convolutional layer corresponding to the encoding convolutional layer. The concatenation is used to compensate for the loss caused by the decoding convolutional layer in restoring the data. When the decoding convolution layer restores the data, the restoration process is performed based on the smallest size, so the details in the data will be lost. Therefore, the present invention compensates for the loss of detail by concatenating the encoding convolutional layers and decoding the convolutional layers. In this embodiment, the decoding layer 221 may be a concatenation of the encoding convolutional layer 217 and the decoding convolutional layer corresponding to the encoding convolutional layer 217 . Decoding layer 222 may be a decoding convolutional layer corresponding to encoding convolutional layer 216 and a concatenation of encoding convolutional layer 216 . Decoding layer 223 may be a decoding convolutional layer corresponding to encoding convolutional layer 215 and a concatenation of encoding convolutional layers 215 . The decoding layer 224 may be a decoding convolutional layer corresponding to the encoding convolutional layer 214 and a concatenation of the encoding convolutional layer 214 . Decoding layer 225 may be a decoding convolutional layer corresponding to encoding convolutional layer 213 and a concatenation of encoding convolutional layer 213 . Decoding layer 226 may only include encoded convolutional layers corresponding to encoded convolutional layers 212 . Decoding layer 227 may only include coded convolutional layers corresponding to coded convolutional layers 211 .

The decoding layer 221 may receive the encoded feature map output by the convolutional encoding layer 218 and perform a deconvolution operation on the encoded feature map to generate a new feature map. The decoding layer 222 may receive the feature map output by the decoding layer 221, and perform a deconvolution operation on the feature map output by the decoding layer 221 to generate a new feature map. In a similar manner, the decoding layer in the decoding convolutional network 220 may receive the feature map output by the previous decoding layer and generate a new feature map according to the received feature map. After deconvolution operations of multiple decoding layers, the decoding layer 227 may perform deconvolution operations on the feature maps output by the decoding layer 226 to generate new feature maps.

The multiple decoding layers in the decoding convolutional network 220 may correspond to different sizes. In other words, the feature maps output by different decoding layers may have different sizes. For example, the size of the feature map output by the decoding layer 221 may be different from the size of the feature map output by the decoding layer 222. The decoding convolutional network 220 can utilize multiple decoding layers of different sizes to extract important features of the original document 30 in time or space at multiple scales.

In one embodiment, the multiple decoding layers in the decoding convolutional network 220 may be moving inverse bottleneck convolutional layers. Taking the decoding layer 221 as an example, the decoding layer 221 may perform a moving inverse bottleneck convolution on the feature map output by the encoding convolutional layer 218 to generate a new feature map. Taking the decoding layer 222 as an example, the decoding layer 222 may perform a moving inverse bottleneck convolution on the feature map output by the decoding layer 221 to generate a new feature map.

The first model 230 may be a kind of neural network. For example, the first model 230 may be a context segmentation network. The contiguous pyramid module 231 of the first model 230 may generate a segmented file corresponding to the original file 30 from the output of one or more decoding layers in the decoding convolutional network 220 . FIG. 4 is a schematic diagram illustrating a process of generating the segmented file 70 by the close pyramid module 231 according to an embodiment of the present invention. Specifically, in the process (a), the contiguous pyramid module 231 can obtain one or more decoding layers in the decoding convolutional network 220 whose distance is closer to that of the contiguous pyramid module 231 (ie, the distance-encoding convolutional network). One or more feature maps output by one or more decoding layers that are closer to the output of 220), wherein the one or more decoding layers may include the decoding layer closest to the close pyramid module 231 (ie: for A decoding layer 227) that produces the output of the decoding convolutional network 220). Next, the close-contact pyramid module 231 may perform convolution operations on the acquired feature maps to generate new feature maps.

In this embodiment, the contiguous pyramid module 231 can obtain the feature map 53 , the feature map 52 and the feature map 51 from the decoding layer 227 , the decoding layer 225 and the decoding layer 224 respectively, wherein the size of the feature map 53 can be larger than that of the feature map 52 , And the size of the feature map 52 may be larger than that of the feature map 51 . The close pyramid module 231 may perform a convolution operation on the feature map 51, the feature map 52, and the feature map 53 to generate the feature map 54, the feature map 55, and the feature map 56, respectively, wherein the feature map 56 may be larger in size than the feature map 55, and Feature map 55 may be larger in size than feature map 54 .

In order to make the size of the feature maps the same, in the process (b), the close pyramid module 231 may upsample the feature maps with smaller size. In this embodiment, the contact pyramid module 231 may upsample the feature map 54 to generate the feature map 57 , wherein the size of the feature map 57 may be the same as the size of the feature map 56 . The contiguous pyramid module 231 may upsample the feature map 55 to generate the feature map 58 , where the size of the feature map 58 may be the same as the size of the feature map 56 .

Next, the close pyramid module 231 may add the respective feature maps of the same size to generate a new feature map. The close pyramid module 231 may concatenate the feature maps generated according to the respective feature maps and the respective feature maps to generate a new feature map. Assuming that the close pyramid module 231 wants to connect N+1 (N is a positive integer) feature maps, the close pyramid module 231 can follow: the feature maps generated according to the respective feature maps, corresponding to the distance from the first model 230 The feature map of the decryption layer at the first distance, the feature map of the decryption layer corresponding to the second distance from the first model 230, ..., the feature map of the decryption layer corresponding to the Nth distance from the first model 230, wherein the A distance may be less than the second distance, and the second distance may be less than the Nth distance. In this embodiment, the close-contact pyramid module 231 may add the feature map 56 , the feature map 57 , and the feature map 58 to generate the feature map 59 . Next, the abutment pyramid module 231 may sequentially connect the feature map 59 , the feature map 56 , the feature map 58 , and the feature map 57 to generate the feature map 5 .

In the process (c), the close pyramid module 231 may input the feature map to an inverted residual block (IRB) to amplify the compensation for the spatial information of the original document. The close-connected pyramid module 231 may perform an atrous convolution (atrous convolution) operation or a separable convolution (S-CONV) operation on the output of the reverse residual block based on different dilation rates to generate multiple feature map. In this embodiment, the close pyramid module 231 can input the feature map 5 to the inverse residual block, and based on the dilation rate 1 (D=1), the dilation rate 2 (D=2), the dilation rate 4 (D= 4) and a dilation rate of 8 (D=8), respectively, perform a dilated convolution operation on the output of the reverse residual block to generate 4 feature maps, which are feature map 61, feature map 62, feature map 63, and feature map 64. That is, feature map 61 corresponds to expansion rate 1, feature map 62 corresponds to expansion rate 2, feature map 63 corresponds to expansion rate 4, and feature map 64 corresponds to expansion rate 8.

In the process (d), the close pyramid module 231 may add the feature maps with the same size to generate a new feature map. The close pyramid module 231 can connect the respective feature maps and the feature maps generated according to the respective feature maps to generate a new feature map. In this embodiment, the close pyramid module 231 may add the feature map 61 , the feature map 62 , the feature map 63 and the feature map 64 to generate the feature map 65 . Next, the close pyramid module 231 may sequentially connect the feature map 61 , the feature map 62 , the feature map 63 , the feature map 64 , and the feature map 65 to generate the feature map 6 . The contiguous pyramid module 231 may perform a convolution operation on the feature map 6 to generate the segmented file 70 . The processor 110 may output the segmented file 70 through the transceiver 130 .

In one embodiment, the first model 230 may further input the segmented document 70 output by the close pyramid module 231 to the extrusion and excitation network 232 to enhance the features of the segmented document 70 . Squeeze and excite network 232 may generate segmented file 40 from segmented file 70 . The processor 110 may output the segmented file 40 through the transceiver 130 .

The second model 240 may be a neural network. For example, the second model 240 may be an edge supervision network. The second model 240 may generate a segmented file corresponding to the original file 30 from the output of one or more decoding layers in the decoding convolutional network 220 . FIG. 5 is a schematic diagram illustrating a process of generating the segmented file 50 by the second model 240 according to an embodiment of the present invention. Specifically, in the process (A), the second model 240 can obtain one or more decoding layers in the decoding convolutional network 220 whose distances are closer to the second model 240 (ie: the distance coding convolutional network 220 ) one or more feature maps output by the one or more decoding layers that are closer to the output end), where the one or more decoding layers may include the decoding layer closest to the second model 240 (ie: used to generate the decoded volume) Decoding layer 227) of the output of product network 220). Next, the second model 240 may perform a convolution operation on the acquired feature maps to generate new feature maps.

In this embodiment, the second model 240 can obtain the feature map 83 , the feature map 82 , and the feature map 81 from the decoding layer 227 , the decoding layer 225 , and the decoding layer 224 , respectively, wherein the size of the feature map 83 can be larger than that of the feature map 82 , and The size of the feature map 82 may be larger than that of the feature map 81 . In one embodiment, feature map 81, feature map 82, and feature map 83 may be the same as feature map 51, feature map 52, and feature map 53, respectively. The second model 240 may perform a convolution operation on the feature map 51 , the feature map 52 , and the feature map 53 to generate the feature map 84 , the feature map 85 , and the feature map 86 , respectively, wherein the feature map 86 may be larger in size than the feature map 85 , and the features Figure 85 may be larger in size than feature map 84 .

In order to make the size of the feature maps the same, in the process (B), the second model 240 may upsample the feature maps with smaller sizes. In this embodiment, the second model 240 may upsample the feature map 58 to generate the feature map 87 , where the feature map 87 may be the same size as the feature map 86 . The second model 240 may upsample the feature map 85 to generate the feature map 88 , where the feature map 88 may be the same size as the feature map 86 .

Next, the second model 240 may concatenate the respective feature maps of the same size to generate a new feature map. Assuming that the second model 240 wants to connect M (M is a positive integer) feature maps, the second model 240 can follow: a feature map corresponding to a decryption layer at a first distance from the second model 240, a feature map corresponding to the second model 240 240 the feature maps of the decryption layer at a second distance, ..., corresponding to the order of the feature maps of the decryption layer at the Mth distance from the second model 240 to connect the M feature maps, where the first distance may be greater than the second distance, and the second distance may be greater than the Mth distance. In this embodiment, the second model 240 can connect the feature map 87 , the feature map 88 and the feature map 86 in sequence to generate the feature map 8 .

In process (C), the second model 240 may perform a convolution operation on the feature map 8 to generate the feature map 50 . The processor 110 may output the feature map 50 through the transceiver 130 .

為對應於第i筆訓練資料和第j個分類的預測結果，並且

為對應於第i筆訓練資料和第j個分類的真值（ground-truth）。處理器110可根據損失函數L來訓練神經網路模型200以調整編碼卷積網路210、解碼卷積網路220、第一模型230及/或第二模型240的超參數，藉以最佳化神經網路模型200的效能。

The loss function L of the neural network model 200 is shown in the following formula, where L1 is the loss function of the first model 230, L2 is the loss function of the second model 240, n is the number of training data, and m is the number of classifications

is the prediction result corresponding to the ith training data and the jth classification, and

is the ground-truth corresponding to the ith training data and the jth classification. The processor 110 can train the neural network model 200 according to the loss function L to adjust the hyperparameters of the encoding convolutional network 210, the decoding convolutional network 220, the first model 230 and/or the second model 240 for optimization Performance of Neural Network Model 200.

6 illustrates a flowchart of a method for file segmentation according to an embodiment of the present invention, wherein the method may be implemented by the electronic device 100 shown in FIG. 1 . In step S601, an original file and a neural network model are obtained, wherein the neural network model includes a first model, wherein the first model is configured to perform: obtaining a first feature map and a second feature map corresponding to a first size of the original file a second feature map of size, where the first size is greater than the second size; performing a first upsampling on the second feature map to generate a third feature map of a third size, where the third size is equal to the first size; connecting the first features map and a third feature map to generate a fourth feature map; input the fourth feature map to a first inverse residual block to generate a first output, and perform a first dilation rate-based first dilated convolution on the first output operate to generate a fifth feature map; input the fourth feature map to a second inverse residual block to generate a second output, and perform a second dilated convolution operation based on a second dilation rate on the second output to generate a sixth a feature map, wherein the second dilation rate is different from the first dilation rate; concatenating the fifth feature map and the sixth feature map to generate a seventh feature map; and performing a first convolution operation on the seventh feature map to generate a segmented file. In step S603, the divided file is output.

To sum up, the neural network model of the present invention can extract the features of the original document through the encoding convolutional network and the decoding convolutional network to generate multiple feature maps. The first model can concatenate multiple feature maps to generate feature maps that contain important features of the original document at multiple scales in time or space. The first model can also increase the number of channels of the feature map through reverse residual block and dilated convolution operations, thereby compensating for the spatial information of the original file. On the other hand, the present invention can train hyperparameters in the neural network model according to the loss functions of the first model and the second model, so that the trained neural network model has better performance. The architecture of the neural network model proposed by the present invention can generate more accurate file segmentation results under the condition of using less computing resources.

100: Electronics 110: Processor 120: Storage Media 200: Neural Network Models 210: Encoding Convolutional Networks 220: Decoding Convolutional Networks 230: First Model 231: Close Pyramid Module 232: Squeeze and Incentivize the Web 240: Second Model 130: Transceiver 30: Original file 211, 212, 213, 214, 215, 216, 217, 218: encoding convolutional layers 221, 222, 223, 224, 225, 226, 227: decoding layer 40, 50, 70: Split file 5, 51, 52, 53, 54, 55, 56, 57, 58, 59, 6, 61, 62, 63, 64, 65, 8, 81, 82, 83, 84, 85, 86, 87, 88: Feature map S601, S603: steps

FIG. 1 is a schematic diagram illustrating an electronic device for file segmentation according to an embodiment of the present invention. FIG. 2 is a schematic diagram illustrating a neural network model according to an embodiment of the present invention. 3 is a schematic diagram illustrating an original document and a processed document according to an embodiment of the present invention. FIG. 4 is a schematic diagram illustrating a process of generating a segmented file by a dense pyramid module according to an embodiment of the present invention. FIG. 5 is a schematic diagram illustrating a process of generating a segmented file by the second model according to an embodiment of the present invention. FIG. 6 is a flowchart illustrating a method for file segmentation according to an embodiment of the present invention.

S601, S603: steps

Claims (10)

An electronic device for file segmentation, comprising: Transceiver, receiving the original file; storage media for storing neural network models; and a processor, coupled to the storage medium and the transceiver, and to access and execute the neural network model, wherein the neural network model includes a first model, wherein the first model is configured to execute: obtaining a first feature map corresponding to a first size of the original document and a second feature map of a second size, wherein the first size is larger than the second size; performing a first upsampling on the second feature map to generate a third feature map of a third size, wherein the third size is equal to the first size; connecting the first feature map and the third feature map to generate a fourth feature map; inputting the fourth feature map to a first inverse residual block to generate a first output, and performing a first dilation rate-based first dilated convolution operation on the first output to generate a fifth feature map; The fourth feature map is input to a second inverse residual block to generate a second output, and a second dilated convolution operation based on a second dilation rate is performed on the second output to generate a sixth feature map, wherein the second expansion rate is different from the first expansion rate; connecting the fifth feature map and the sixth feature map to produce a seventh feature map; and performing a first convolution operation on the seventh feature map to produce a segmented file, wherein The processor outputs the segmented file through the transceiver. The electronic device of claim 1, wherein the neural network model further comprises a second model, wherein the second model is configured to perform: performing a second upsampling on the second feature map to generate an eighth feature map of a fourth size, wherein the fourth size is equal to the first size; connecting the first feature map and the eighth feature map to generate a ninth feature map; and A second convolution operation is performed on the ninth feature map to generate an output feature map. The electronic device of claim 2, wherein the first model corresponds to a first loss function, wherein the second model corresponds to a second loss function, wherein the processor combines the first loss function and the adding the second loss functions to generate a third loss function, wherein the processor trains the first model and the second model according to the third loss function. The electronic device of claim 1, wherein the neural network model further comprises a coding convolutional network, wherein the coding convolutional network comprises a first coding convolutional layer and a second coding convolutional layer, wherein the coding Convolutional Networks are configured to perform: generating a first encoded feature map from the original file and the first encoded convolutional layer; and A second encoded feature map is generated from the first encoded feature map and the second encoded convolutional layer. The electronic device of claim 4, wherein the neural network model further comprises a decoding convolutional network, wherein the decoding convolutional network comprises a first decoding layer and a second decoding layer, wherein the first decoding The layers include the second encoded convolutional layer and a decoded convolutional layer corresponding to the second encoded convolutional layer, wherein the decoded convolutional network is configured to perform: generating the second feature map from the second encoding feature map and the first decoding layer; and The first feature map is generated from the second feature map and the second decoding layer. The electronic device of claim 1, wherein the first model is further configured to perform: adding the first feature map and the third feature map to generate a tenth feature map; and The tenth feature map, the first feature map, and the third feature map are concatenated to generate the fourth feature map. The electronic device of claim 1, wherein the first model is further configured to perform: adding the fifth feature map and the sixth feature map to generate an eleventh feature map; and The fifth feature map, the sixth feature map, and the eleventh feature map are concatenated to generate the seventh feature map. The electronic device of claim 1, wherein the first model is further configured to perform: performing the first convolution operation on the seventh feature map to generate a twelfth feature map; and The twelfth feature map is input to a squeeze and excitation network to generate the segmented file. The electronic device of claim 4, wherein the first encoding convolutional layer performs a moving reverse bottleneck convolution on the original document to generate the first encoding feature map. A method for file segmentation, comprising: Obtain the original document and a neural network model, wherein the neural network model includes a first model, wherein the first model is configured to perform: obtaining a first feature map corresponding to a first size of the original document and a second feature map of a second size, wherein the first size is larger than the second size; performing a first upsampling on the second feature map to generate a third feature map of a third size, wherein the third size is equal to the first size; connecting the first feature map and the third feature map to generate a fourth feature map; inputting the fourth feature map to a first inverse residual block to generate a first output, and performing a first dilation rate-based first dilated convolution operation on the first output to generate a fifth feature map; The fourth feature map is input to a second inverse residual block to generate a second output, and a second dilated convolution operation based on a second dilation rate is performed on the second output to generate a sixth feature map, wherein the second expansion rate is different from the first expansion rate; connecting the fifth feature map and the sixth feature map to produce a seventh feature map; and performing a first convolution operation on the seventh feature map to generate a segmented file; and The split file is output.

Images