System and method for adjusting neural network input data

Abstract

The present invention provides a system and method for adjusting input data input to a decision neural network, the system comprising: a segmented data neural network device and a data processing device. The segmented data neural network device receives input data and segments the input data into a plurality of partitioned data including first partitioned data and second partitioned data; the data processing device is coupled to the segmented data neural network device to receive each partitioned data, and processes the first partitioned data and the second partitioned data in different ways when processing these partitioned data so that different changes occur in the first partitioned data and the second partitioned data. The decision neural network is electrically coupled to the data processing device to use the processed partitioned data as input data. In this way, the neural network can change the final output result.

Description

Technical Field

[0001] The present invention relates to a system for assisting a neural network, and more particularly to a system and method for adjusting neural network input data. Background Art

[0002] Neural Network is a computer application that performs calculations through a large number of artificial neurons. In most cases, neural networks can gradually change their internal structure based on external information, so they have the function of learning to some extent. After the neural network is built, it can train itself by accumulating external information and gradually improve its structure, and can provide quite accurate judgment results after training.

[0003] When a neural network can provide a fairly accurate judgment result, it means that the same neural network can only produce the same result for the same external input condition. From another perspective, if the neural network needs to produce other results for the same external input condition, then the neural network must be retrained to achieve the goal. However, each process of training itself by accumulating external information must take a lot of time, so when facing an operating program that may frequently change the judgment conditions, introducing a neural network may be a thankless task.

[0004] For example, if the packaging seams of product A only require the gap between adjacent seams to be less than 1mm to meet the requirements, while the packaging seams of product B require the gap between adjacent seams to be less than 0.5mm to meet the requirements, then the visual neural network used to inspect the packaging seams of product A and the visual neural network used to inspect the packaging seams of product B cannot be shared. In this case, the existing technology can only solve the problem by establishing two sets of visual neural networks to correspond to the seam inspection of two different products respectively. Once other products with different requirements for packaging seams appear, the problem can only be solved by increasing the number of visual neural networks or retraining the visual neural networks that are no longer used.

[0005] From this we can see that although neural networks can provide accurate judgment results, they are not suitable as a means of solving problems in certain special environments because they can only provide accurate judgment results. Summary of the invention

[0006] In view of this, the description of the present invention proposes a system and method for adjusting neural network input data, which changes the data content originally input to the neural network so that the neural network changes the final output result.

[0007] From one perspective, the present invention proposes a system for adjusting neural network input data, which is suitable for adjusting input data input to a decision neural network and includes: a data segmentation neural network device and a data processing device. The data segmentation neural network device receives input data and segments the input data into a plurality of partition data including first partition data and second partition data; the data processing device is coupled to the data segmentation neural network device to receive each partition data, and when processing these partition data, processes the first partition data and the second partition data in different ways to make different changes to the first partition data and the second partition data, and collects the changed first partition data and the changed second partition data as modified input data. Among them, the decision neural network is coupled to the data processing device to use the modified input data as input data, and the decision neural network makes a judgment based on the relative relationship between the first partition data and the second partition data when generating a result, and the data processing device changes the relative relationship between the first partition data and the second partition data.

[0008] In one embodiment, the aforementioned input data is an image.

[0009] In one embodiment, each partition data is a portion of the aforementioned image, and after being processed by the data processing device, the distance between the first portion of the image represented by the first partition data and the second portion of the image represented by the second partition data changes. Further, the decision neural network determines the output result according to the distance between the first portion of the image and the second portion of the image.

[0010] In another embodiment, the data processing device causes the first partition data and the second partition data to produce multiple changes in relative relationship to different degrees to be combined into a corresponding multiple different modified input data, wherein these different modified input data are input into multiple decision neural networks to generate corresponding multiple different results, and these decision neural networks generate results based on the relative relationship between the first partition data and the second partition data.

[0011] In one embodiment, the first partial image represented by the first partition data and the second partial image represented by the second partition data do not overlap with each other.

[0012] In one embodiment, the data segmentation neural network device uses a semantic analysis neural network to segment the input data to generate the aforementioned plurality of partitioned data.

[0013] In one embodiment, the decision neural network first receives input data as input to generate the result. When the result does not meet the standard result, the system receives the same input data again to generate modified input data, and the decision neural network receives the modified input data as input again to regenerate the result.

[0014] In one embodiment, the data processing device determines the relative relationship between the first partition data and the second partition data to be changed according to the result generated by inputting the input data into the decision neural network.

[0015] From another perspective, the present invention proposes a method for adjusting neural network input data, which is suitable for adjusting the input data input to the decision neural network to generate a result. The method includes: receiving input data; dividing the input data into a plurality of partition data including first partition data and second partition data; processing the first partition data and the second partition data in different ways so that the first partition data and the second partition data produce different changes and change the relative relationship between the first partition data and the second partition data; combining the changed first partition data and the changed second partition data as modified input data; and using the modified input data as the input of the decision neural network, wherein the decision neural network makes a judgment based on the relative relationship between the first partition data and the second partition data to generate a result.

[0016] In one embodiment, the aforementioned input data is an image.

[0017] In one embodiment, each partition data is a portion of an image, and after processing the first partition data and the second partition data in different ways, the distance between the first portion of the image represented by the first partition data and the second portion of the image represented by the second partition data changes. Further, the decision neural network determines the output result according to the distance between the first portion of the image and the second portion of the image.

[0018] In one embodiment, when the first partition data and the second partition data are processed in different ways, the first partition data and the second partition data are caused to produce multiple different degrees of relative relationship changes to be aggregated into corresponding multiple different modified input data, wherein these different modified input data are input into multiple decision neural networks to generate corresponding multiple different results, and these decision neural networks make judgments based on the above-mentioned relative relationship between the first partition data and the second partition data when generating the results.

[0019] In one embodiment, the first partial image represented by the first partition data and the second partial image represented by the second partition data do not overlap with each other.

[0020] In one embodiment, the method uses a semantic analysis neural network to segment input data to generate the aforementioned partition data.

[0021] In one embodiment, before processing the first partition data and the second partition data in different ways so as to produce different changes in the first partition data and the second partition data and thus change the relative relationship between the first partition data and the second partition data, it also includes: inputting input data into a decision neural network to generate a result; when the generated result meets the standard result, outputting the result, and not executing the step and subsequent steps of processing the first partition data and the second partition data in different ways so as to produce different changes in the first partition data and the second partition data and thus change the relative relationship between the first partition data and the second partition data; and when the generated result does not meet the standard result, executing the step and subsequent steps of processing the first partition data and the second partition data in different ways so as to produce different changes in the first partition data and the second partition data and thus change the relative relationship between the first partition data and the second partition data.

[0022] By adopting the above technology, the present invention can change the data originally to be provided to the decision neural network. Before providing the input data to the decision neural network, the system provided by the present invention can first divide the input data into multiple partition data and change the relationship between the multiple partition data. These partition data with changed relationships will be further input into the decision neural network. Therefore, after the above processing, the input data provided to the decision neural network has changed, and the decision neural network generates results based on the changed input data. Therefore, as long as the parameters of the data processing device are appropriately adjusted, the purpose of ultimately changing the output result of the decision neural network can be achieved. Since the time required to adjust the parameters of the data processing device is obviously lower than the time required to retrain the neural network, the technology provided by the present invention can reduce the time cost required when using the neural network in certain special environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 FIG. 4 is a circuit block diagram of a system for adjusting neural network input data according to an embodiment of the present invention.

[0024] Figure 2 FIG. 4 is a flowchart of an operation of a system for adjusting neural network input data according to an embodiment of the present invention.

[0025] Figure 3 is a schematic diagram of input data according to an embodiment of the present invention.

[0026] Figure 4 is a schematic diagram of input data according to an embodiment of the present invention.

[0027] Figure 5A FIG. 4 is a schematic diagram of changed first partition data according to an embodiment of the present invention.

[0028] Figure 5BFIG. 4 is a schematic diagram of modified input data according to an embodiment of the present invention.

[0029] Among them, the simple description of the symbols in the accompanying drawings is as follows:

[0030] 10: system for adjusting neural network input data; 15: decision neural network; 100: segmentation data neural network device; 110: data processing device; 300: image; B1, B2, B3: sub-blocks; d1, d2: distances; INM: input data; INM1: first partition data; INM2: second partition data; INM': modified input data; P1~P9, P1'~P9', Q1~Q 12 , R1~R 18 : position; S1, S2, S3: line segments; S200~S206: implementation steps of an embodiment of the present invention. DETAILED DESCRIPTION

[0031] Please refer to Figure 1 , which is a circuit block diagram of a system for adjusting neural network input data according to an embodiment of the present invention. Figure 1 As shown, the system for adjusting neural network input data (hereinafter referred to as the system) 10 includes a segmentation data neural network device 100 and a data processing device 110, wherein the segmentation data neural network device 100 receives the input data INM originally to be input to the decision neural network 15, and the modified input data INM' output by the data processing device 110 is provided to the decision neural network 15 as data to replace the input data INM. Finally, the decision neural network 15 generates a final decision result according to the received modified input data INM'. In one embodiment, the segmentation data neural network device 100 and the data processing device 110 can each include a processor to execute the corresponding program code to realize their respective functions; in another embodiment, a processor included in the system 10 can also execute the program codes corresponding to the segmentation data neural network device 100 and the data processing device 110 to realize their functions respectively. In one embodiment, the decision neural network 15 can be a convolutional neural network Xception, but those skilled in the art can know that there are many existing convolutional neural networks that can realize decision neural networks, and the present invention is not limited thereto.

[0032] Please refer to Figure 1 and Figure 2 ,in, Figure 21 is an operation flow chart of a system for adjusting neural network input data according to an embodiment of the present invention. In this embodiment, the segmented data neural network device 100 first receives input data INM to be provided to the decision neural network 15 (step S200). The input data INM can be data of any dimension, such as data composed of sequence text or images. The following description will be carried out using images as input data INM, but the actual application is not limited to this.

[0033] Please also refer to Figure 3 , which is a schematic diagram of input data according to an embodiment of the present invention. In this embodiment, image 300 is the input data (equivalent to the above-mentioned input data INM) originally to be provided to the decision neural network 15, which includes line segment S1, line segment S2 and line segment S3. In image 300, the distance between line segment S1 and line segment S2 is d1, and the distance between line segment S1 and line segment S3 or line segment S2 and line segment S3 is d2. After the segmentation data neural network device 100 receives the image 300 through step S200, the segmentation data neural network device 100 will capture the features of the image 300 and perform segmentation according to the pre-trained neural network model. It is worth mentioning that the segmentation data neural network device 100 should be trained to be able to segment the image 300 according to the decision basis of the decision neural network 15. For example, when the decision neural network 15 determines the decision result according to the distance between the line segment S1 and the line segment S2, the segmentation data neural network device 100 must be able to segment the sub-block B1 including the line segment S1 and the sub-block B2 including the line segment S2 from the image 300; when the decision neural network 15 determines the decision result according to the distance between the line segment S1 and the line segment S3, the segmentation data neural network device 100 must be able to segment the sub-block B1 including the line segment S1 and the sub-block B3 including the line segment S3 from the image 300. In general, the segmentation data neural network device 100 segments the image 300 and the data matrix (i.e., its pixel matrix) behind it according to the judgment object (the line segment S1 and the line segment S2 or the line segment S1 and the line segment S3 in the above example) used by the decision neural network 15 when making a decision. In different practical applications, the segmentation data neural network device 100 can use various methods to identify the above-mentioned sub-blocks, such as semantic analysis (semantic analysis) neural network methods, to train and improve its ability to correctly cut images. In one embodiment, the segmentation data neural network device 100 can be implemented by a convolutional neural network U-Net. U-Net can be trained to perform semantic analysis based on the basis of the decision result generated by the decision neural network 15 to appropriately segment the image 300, but the present invention is not limited to this. Those skilled in the art can know that there are many existing convolutional neural networks that can implement semantic analysis, which will not be repeated here.

[0034] While cutting out the sub-blocks B1, B2 or B3, the data segmentation neural network device 100 can also divide the corresponding data content of the sub-blocks B1, B2 or B3 in the input data INM into corresponding partition data. In one embodiment, the data segmentation neural network device 100 divides the partition data by establishing corresponding masks. Figure 4 , assuming that the original input data INM is an N*M data matrix (e.g., an image with a resolution of N*M), the data content corresponding to the sub-block B1 in the input data INM is stored in Figure 4 The positions marked as P1 to P9 (e.g., pixel positions in the image) and the corresponding data contents of the sub-block B2 in the input data INM are stored in Figure 4 Indicated as Q1~Q 12 The data content corresponding to the sub-block B3 in the input data INM is stored in Figure 4 Indicated as R1~R 18 In this case, the segmentation data neural network device 100 can create three masks of size N*M. One of the three masks (hereinafter referred to as the first mask) will be Figure 4 The values ​​of the positions marked as P1 to P9 in the figure are set to 1, and the values ​​of other positions are set to 0; another mask (hereinafter referred to as the second mask) is set in Figure 4 Indicated as Q1~Q 12 The value of the position is set to 1, and the values ​​of other positions are set to 0; the last mask (hereinafter referred to as the third mask) will be Figure 4 Indicated as R1~R 18 The value of the position is set to 1, and the value of other positions is set to 0.

[0035] Next, the data segmentation neural network device 100 can use the three masks to perform an AND operation with the input data INM and generate corresponding partition data. That is, the data segmentation neural network device 100 uses the first mask to perform an AND operation with the input data INM to generate a corresponding partition data. Figure 4 Except for the values ​​of the positions marked as P1 to P9 in the figure, the values ​​of other positions will become 0 due to the first mask. Similarly, the segmentation data neural network device 100 uses the second mask to perform an AND operation with the input data INM to generate a corresponding partition data. Figure 4 Indicated as Q1~Q 12The value of the position remains unchanged, and the values ​​of other positions will become 0 due to the relationship of the second mask; the segmentation data neural network device 100 uses the third mask and the input data INM to perform an "AND" operation to generate a corresponding partition data. The content of this partition data is Figure 4 Indicated as R1~R 18 Except for the value of the position that remains unchanged, the values ​​of other positions will become 0 due to the third mask.

[0036] It should be noted that when the decision neural network 15 determines the decision result according to the relative relationship between the line segment S1 and the line segment S2, the segmentation data neural network device 100 only needs to segment the sub-block B1 including the line segment S1 and the sub-block B2 including the line segment S2, and only needs to generate the first mask and the second mask to calculate and obtain two corresponding partition data. Similarly, when the decision neural network 15 determines the decision result according to the relative relationship between the line segment S1 and the line segment S3, the segmentation data neural network device 100 only needs to segment the sub-block B1 including the line segment S1 and the sub-block B3 including the line segment S3, and only needs to generate the first mask and the third mask to calculate and obtain two corresponding partition data. Only when the decision neural network 15 determines the decision result according to the relative relationship between the line segment S1, the line segment S2 and the line segment S3, the segmentation data neural network device 100 needs to simultaneously generate the first, second and third masks and calculate and obtain three corresponding partition data.

[0037] The following description assumes that the decision neural network 15 generates a decision result according to the relative relationship between line segment S1 and line segment S2, but technicians in this field should know that the relevant content can also be extrapolated to the situation where the decision result is generated according to the relative relationship between line segment S1 and line segment S3 or according to the relative relationship between line segment S1, line segment S2 and line segment S3.

[0038] Please come back Figure 1 and Figure 2, when the decision neural network 15 generates a decision result according to the relative relationship between the line segment S1 and the line segment S2, the segmentation data neural network device 100 will segment the sub-block B1 and the sub-block B2 from the image 300 after receiving the image 300 (step S200), and further obtain the corresponding mask and the corresponding partition data (step S202). The partition data corresponding to the sub-block B1 (hereinafter referred to as the first partition data) INM1 and the partition data corresponding to the sub-block B2 (hereinafter referred to as the second partition data) INM2 will be transmitted to the data processing device 110 in any possible form. For example, the first partition data INM1 and the second partition data INM2 can be transmitted separately, or new data formed by combining the first partition data INM1 and the second partition data INM2 through an OR operation can be transmitted.

[0039] After receiving the first partition data INM1 and the second partition data INM2, the data processing device 110 will start to process the first partition data INM1 and the second partition data INM2 accordingly. It should be noted that the data processing performed by the data processing device 110 on the first partition data INM1 and the second partition data INM2 should be related to the relative relationship according to which the decision neural network 15 generates the decision result. For example, when the decision neural network 15 generates the decision result according to the distance between the first partition data INM1 and the second partition data INM2 (that is, the distance d1 between the line segment S1 and the line segment S2 on the image 300), the processing operation to be performed by the data processing device 110 may include: moving the position of the non-0 data of the first partition data INM1, moving the position of the non-0 data of the second partition data INM2, or moving the position of the non-0 data of the first partition data INM1 and the second partition data INM2 at the same time, so as to change the value of the distance d1. In another example, when the decision neural network 15 generates a decision result according to the brightness difference between the first partition data INM1 and the second partition data INM2, the processing operation to be performed by the data processing device 110 may include: changing the value of the non-0 data of the first partition data INM1, changing the value of the non-0 data of the second partition data INM2, or changing the value of the non-0 data of the first partition data INM1 and the second partition data INM2 at the same time, so as to thereby change the brightness difference between the first partition data INM1 and the second partition data INM2. In one embodiment, the decision result generated by the decision neural network 15 may be one of multiple possible results, wherein the multiple possible results may be generated according to at least one relative relationship between each partition data, or may be unrelated to the relative relationship between each partition data. For example, the multiple possible results may include: the distance between line segment S1 and line segment S2 is too large (generated according to the distance between the partition data), the distance between line segment S2 and line segment S3 is too small (generated according to the distance between the partition data), the brightness difference between line segment S1 and line segment S2 is too small (generated according to the brightness difference between the partition data), damage appears in the image (unrelated to the relative relationship between the partition data), line segments appear in the image (unrelated to the relative relationship between the partition data), meet the standard (i.e., do not include any results that do not meet the standard), etc. In other embodiments, the processing operation performed by the data processing device 110 may also be other operations that change each partition data, such as multiplying different partition data by different magnification / reduction ratios, or performing different filtering on different partition data, etc.

[0040] In general, the data processing device 110 changes the relative relationship between the first partition data INM1 and the second partition data INM2 by adjusting the contents thereof (step S204). The changed first partition data INM1' and the changed second partition data INM2' are integrated into a new modified input data INM' by an "OR" operation and provided to the decision neural network 15 (step S206). For example, when the data processing device 110 chooses to move the non-zero data of the first partition data INM1 to the left by one grid in order to change the distance between the first partition data INM1 and the second partition data INM2 (equivalent to changing the distance between the line segment S1 and the line segment S2), the content of the changed first partition data INM1' will be as follows. Figure 5A As shown in the figure, the values ​​originally recorded at the positions marked as P1 to P9 will be respectively moved one grid to the left and recorded at the positions marked as P1' to P9'. Moreover, the content values ​​that exceed the range of the N*M matrix due to the left shift are discarded, and the content values ​​that need to be supplemented on the right side of the matrix due to the left shift are set to 0. In this case, when the data processing device 110 integrates the changed first partition data INM1' and the changed second partition data INM2' (in fact, there is no modification in this embodiment, so it is equivalent to the second partition data INM2) into the modified input data INM' through an "OR" operation, the modified input data INM' will include the following. Figure 5B The contents shown in the figure include the following: 12 All values ​​outside of the position are 0.

[0041] After the above operations, the modified input data INM' output by the data processing device 110 is also a matrix with a size of N*M like the original input data INM, so the decision neural network 15 can directly use the modified input data INM' as its input and generate the final decision result according to the content of the modified input data INM'. It can be seen that the adjustment made by the data processing device 110 to the data of each partition will affect the decision result generated by the decision neural network 15. Compared with the original input data INM, the content of the modified input data INM' is 0 except for the positions set to 1 in each mask; while in the original input data INM, the positions other than the positions set to 1 in each mask may originally be non-0 values. That is, in the original input data INM, other positions other than the partition data (e.g., INM1 and INM2) that the decision neural network 15 uses to determine its decision result may originally have values ​​other than 0, but in the modified input data INM', the values ​​of other positions other than the partition data (e.g., INM1' and INM2') that the decision neural network 15 uses to determine its decision result are all 0. Therefore, in order to enable the decision neural network 15 to determine the decision result more accurately, in one embodiment, when training the decision neural network 15, the training data used for training may include, in addition to a plurality of original image data, corresponding plurality of mask image data in which the pixel values ​​of other positions other than the partition data that the decision neural network 15 determines the decision result are all set to 0 in the above-mentioned original image data. In other words, the training data used for training the decision neural network 15 may include the above-mentioned plurality of original image data and the above-mentioned plurality of mask image data, so that no matter whether the modified input data INM' or the original input data INM is input to the decision neural network 15, the decision neural network 15 can accurately determine the decision result.

[0042] Still taking the packaging seam of the product as an example, assuming that the decision neural network 15 is originally trained to identify the product as being in a properly packaged state only when the distance d1 between the line segment S1 and the line segment S2 is not greater than 1 mm, once the standard is modified to identify the product as being in a properly packaged state only when the distance d1 between the line segment S1 and the line segment S2 is not greater than 2 mm, the engineer can use the data processing device 110 to adjust the distance between the line segment S1 and the line segment S2 so that when the distance d1 in the image 300 is 2 mm, it can be reduced to meet the recognition standard of the decision neural network 15 due to processing (for example, reducing the distance between the line segment S1 and the line segment S2 by 1 mm). In other words, although the distance d1 in the image 300 is 2 mm, in the modified input data (equivalent to the adjusted image) output by the data processing device 110 to the decision neural network 15, the distance d1 that was originally 2 mm can be changed to 1 mm; in this way, the decision neural network 15 will determine that the packaging seam of this product meets the standard because it receives an image (or data) with a distance d1 of 1 mm. In one embodiment, at the beginning, the original input data INM (such as image 300) is directly input into the decision neural network 15 to generate a decision result. If the decision result meets the standard, it is completed (for example, when the distance d1 is 0.9 mm); if the decision result is that the distance between the line segment S1 and the line segment S2 is too large and does not meet the standard (for example, when the distance d1 is 1.8 mm), the original input data INM is input into the system 10 and the data processing device 110 adjusts the distance between the line segment S1 and the line segment S2 (for example, the distance d1 is reduced by 1 mm and becomes 0.8 mm) and then enters the decision neural network 15 to regenerate the decision result. In this way, the situation where the distance d1 between the line segment S1 and the line segment S2 is not greater than 2 mm can be deemed to meet the standard. In another embodiment, if the decision result generated by the decision neural network 15 is other types of non-standard results that are not related to the line segment distance (for example, the bag surface of the product is damaged or the packaging seam of the product is broken, etc.), the above-mentioned input system 10 and the data processing device 110 will not be used to adjust the distance between the line segments and then regenerate the decision result. Instead, the non-standard result is directly reported.In other embodiments, the adjustment method required by the data processing device 110 may be determined based on different decision results generated by the original input data INM input into the decision neural network 15. For example, if the decision result is that the distance between the first partition data INM1 and the second partition data INM2 is too large and does not meet the standard result, the data processing device 110 adjusts the distance between the partition data and re-inputs the data into the decision neural network 15 to generate a decision result; if the decision result is that the brightness difference between the first partition data INM1 and the second partition data INM2 is too small and does not meet the standard result, the data processing device 110 adjusts the brightness difference between the partition data and re-inputs the data into the decision neural network 15 to generate a decision result.

[0043] In another embodiment, the data processing device 110 may adjust the partition data by performing multiple adjustments of different degrees at the same time to generate multiple different modified input data INM', and then input each different modified input data INM' into multiple parallel and identical decision neural networks 15 to regenerate multiple different decision results, so as to reduce the need to manually change the adjustment degree and regenerate the decision results repeatedly, thereby improving efficiency. Taking the packaging suture of the above-mentioned product as an example, assuming that the decision neural network 15 was originally trained to make the decision result meet the standard only when the distance d2 between the line segment S2 and the line segment S3 is not less than the minimum standard distance of 2mm, and later it is necessary to reduce the above minimum standard distance (for example, to 1.5mm or 1.2mm, etc.), in this embodiment, at the beginning, the original input data INM (for example, image 300) can be directly input into the decision neural network 15 to generate a decision result. If the decision result is that the distance d2 between the line segment S2 and the line segment S3 is too small, the decision result meets the standard. The original input data INM is input into the system 10 and the data processing device 110 simultaneously generates a plurality of different modified input data INM' corresponding to different adjustment distances between the line segment S2 and the line segment S3 (for example, the distance d2 is increased by 0.5mm, 1mm, 1.5mm, etc.), and then each different modified input data INM' is simultaneously input into a plurality of parallel and identical decision neural networks 15 to regenerate a plurality of corresponding different decision results. As long as at least one of the plurality of decision results meets the standard, it is considered to meet the standard. In another possible embodiment, the data processing device 110 may also perform a plurality of different degrees of adjustment in sequence when adjusting the data of each partition, and the plurality of different modified input data INM' may also be sequentially input into one or more decision neural networks 15 to sequentially generate a plurality of corresponding different decision results.

[0044] As can be seen from the above description, the technology provided by the present invention can change the data originally to be provided to the decision neural network. Before providing the input data to the decision neural network, the system can first divide the input data into multiple partition data and change the relationship between the multiple partition data. These partition data with changed relationships will be further input into the decision neural network. Therefore, after the above processing, the input data provided to the decision neural network has changed. Since the decision neural network generates results based on the changed input data, the purpose of ultimately changing the output results of the decision neural network can be achieved by appropriately adjusting the parameters of the data processing device. Since the time required to adjust the parameters of the data processing device is obviously lower than the time required to retrain the neural network, the technology provided by the present invention can reduce the time cost required to use the neural network in certain special environments.

[0045] The above description is only a preferred embodiment of the present invention, but it is not intended to limit the scope of the present invention. Any person familiar with this technology can make further improvements and changes on this basis without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be based on the scope defined by the claims of this application.

Claims

1. A system for adjusting input data of a neural network, adapted to adjust input data input to a decision neural network to produce a result, characterized in that: include: A data segmentation neural network device receives the input data and segments the input data to generate a plurality of partition data including first partition data and second partition data; as well as a data processing device coupled to the segmented data neural network device to receive the plurality of partitioned data, wherein the data processing device processes the first partitioned data and the second partitioned data in different ways when processing the plurality of partitioned data so as to produce different changes to the first partitioned data and the second partitioned data, and collects the changed first partitioned data and the changed second partitioned data as modified input data; In which, the decision neural network is coupled to the data processing device to receive the modified input data as the input of the decision neural network, and the decision neural network generates the result based on the relative relationship between the first partition data and the second partition data, and the data processing device changes the relative relationship between the first partition data and the second partition data by causing different changes to the first partition data and the second partition data.

2. The system of claim 1, wherein the input data is an image.

3. A system as described in claim 2, wherein each of the multiple partition data is a portion of the image, and after being processed by the data processing device, the distance between the first portion of the image represented by the first partition data and the second portion of the image represented by the second partition data changes.

4. The system of claim 3, wherein the decision neural network determines an output result according to a distance between the first partial image and the second partial image.

5. A system as described in claim 1, wherein the data processing device causes the first partition data and the second partition data to produce a plurality of different degrees of change in the relative relationship to be combined into a corresponding plurality of different modified input data, wherein the plurality of different modified input data are input into a plurality of decision neural networks to generate a corresponding plurality of different results, and the plurality of decision neural networks generate results based on the relative relationship between the first partition data and the second partition data. 6 . The system of claim 1 , wherein a first portion of the image represented by the first partition data and a second portion of the image represented by the second partition data do not overlap each other.

7. The system as claimed in claim 1, wherein the data segmentation neural network device uses a semantic analysis neural network to segment the input data to generate the plurality of partitioned data.

8. The system of claim 1, wherein the decision neural network first receives the input data as input to generate a result, and when the result does not meet the standard result, the system receives the input data again to generate the modified input data, and the decision neural network receives the modified input data as input again to regenerate the result.

9. The system as claimed in claim 8, wherein the data processing device determines the relative relationship between the first partition data and the second partition data to be changed according to the result generated by inputting the input data into the decision neural network.

10. A method for adjusting input data of a neural network, suitable for adjusting input data input to a decision neural network to produce a result, characterized in that: include: receiving the input data; Splitting the input data into a plurality of partition data including first partition data and second partition data; Processing the first partition data and the second partition data in different ways to make the first partition data and the second partition data produce different changes and change the relative relationship between the first partition data and the second partition data; The changed first partition data and the changed second partition data are combined as modified input data; as well as The modified input data is used as the input of the decision neural network. The decision neural network generates the result by making a judgment based on the relative relationship between the first partition data and the second partition data. The method of claim 10 , wherein the input data is an image.

12. The method of claim 11, wherein each of the plurality of partition data is a portion of the image, and after the first partition data and the second partition data are processed in different ways, a distance between a first portion of the image represented by the first partition data and a second portion of the image represented by the second partition data changes.

13. The method of claim 12, wherein the decision neural network determines the output result according to the distance between the first partial image and the second partial image.

14. The method of claim 10, wherein when the first partition data and the second partition data are processed in different ways, the first partition data and the second partition data are caused to produce a plurality of different degrees of change in the relative relationship to be aggregated into a corresponding plurality of different modified input data, wherein the plurality of different modified input data are input into a plurality of decision neural networks to produce a corresponding plurality of different results, and the plurality of decision neural networks make judgments based on the relative relationship between the first partition data and the second partition data when generating the results. 15 . The method of claim 10 , wherein the first partial image represented by the first partition data and the second partial image represented by the second partition data do not overlap each other.

16. The method of claim 10, wherein a semantic analysis neural network is used to segment the input data to generate the plurality of partition data.

17. The method of claim 10, wherein before processing the first partition data and the second partition data in different ways to make the first partition data and the second partition data produce different changes and change the relative relationship between the first partition data and the second partition data, the method further comprises: Inputting the input data into the decision neural network to generate the result; When the result meets the standard result, the result is outputted, and the step of processing the first partition data and the second partition data in different ways to make the first partition data and the second partition data produce different changes to change the relative relationship between the first partition data and the second partition data and subsequent steps are not performed; and When the result does not meet the standard result, a step and subsequent steps of processing the first partition data and the second partition data in different ways to make the first partition data and the second partition data produce different changes to change the relative relationship between the first partition data and the second partition data are performed.

18. The method of claim 17, further comprising: The relative relationship between the first partition data and the second partition data to be changed is determined according to the result generated by inputting the input data into the decision neural network.

Images