TW591545B - Method and system for recognizing number by neural network - Google Patents

Abstract

The present invention provides a method and system for recognizing a number in a specific image by a neural network. The specific image includes a portion showing the number. The present invention first retrieves the specific image and transforms it to a gray-scaled image. Then, the present invention positions the portion in accordance with the gray-scaled image and transforms it into a normalization array. After inputting the normalization array to the neural network for further calculation, several values obtained. The present invention compares those values and recognizes the desired number.

Description

591545

FIELD OF THE INVENTION A numerical identification method of a neural network and a target value, and more particularly, a method and system for identifying table readings, for a set of target values in an image. The present invention relates to a utilization system for identifying a specific image, and a method for recognizing a reading on a meter device using a neural network-like instrument. BACKGROUND OF THE INVENTION Current energy companies use the amount of users to further The calculation has increased rapidly, and it is accurate by manual meter reading. In addition, disputes between the company and users may also occur. Manual meter reading to measure various expenses. However, the energy cost caused by the labor cost formula is rather uneconomical and not enough for manual meter reading. At the same time, with the rapid growth of the population, gas is used when many users may share the same gas pipeline. Demolition issues. It is an inevitable problem how to facilitate the sharing and billing of accounts for shared pipeline users and other value-added services. '〃 However, because the manual meter reading method currently used cannot solve the above problems, a practical energy management system is urgently needed to change the user billing method to reduce the cost of energy management. Brief description of the invention As mentioned above, the object of the present invention is to provide a method for utilizing nerve-like

591545

= Road numerical identification method and system for identifying a target value in a specific image. The specific image includes a partial image including a target value. In the method of the present invention, the specific image is first converted into a gray scale format, and the partial image image containing the target value is converted into a normalized image array. After a type of neural network is used to calculate the score, according to the set of analysis values, the first interpretation is performed. Capture this specific image. Then, _ 'and according to the specific image after conversion,. Then, the part containing the target value is converted and the normalized image array is input to analysis, and a set of analysis values is obtained. The most desired target value. In addition, the system of the present invention includes an acquisition device, a positioning device, a normalization device, a neural network, and a disruption. The capture device captures this specific video, mouth = ^ grayscale format. The positioning device outputs a part of the image including the target value according to the position. The normalized image is divided into a normalized image array. Then ^ 1 standard value part to a type of neural network to perform Ye Zengbayu seeing the image array f set according to this set of analysis values, read out the invention to be described in detail. The present invention provides a utilization type of electricity meter, and various industries Users can use independent or decentralized methods to identify gas and meter information via network technology. With this hair delivery system ’, such as a personal computer,

591545 V. Description of the invention (3) Brain, wired and wireless local area network, Internet, personal digital assistant (PDA), mobile phone capable of computing, information appliances with computing power, computers used in cars, boats or aircraft, and separate Or a combination of one circuit or circuit, using one or several neural networks to identify each type ^. 4 Miles In addition, the present invention can further utilize long-distance wired or profitable transmission to reduce manual meter reading, manual filing, manual charging, and artificial < rate and accuracy. However, the method for recognizing the numerical value of the image in the present invention can not only greatly reduce the number of transmissions, but also be beneficial to the subsequent database use and the use of 1 spade and storage of shellfish, meanwhile, bamboo application and shellfish value-added. The present invention uses the Λ-name mouth-pull neural network technology to change the birth rate, and the recognition speed and accuracy of the traditional subtractive method of JL ancient @ t 文 α will have greater stability and adaptability. . X and 槚 The first figure A shows the value ^ < column to which the present invention is applied. Here, the present invention can be applied to the first real value which is distinguished as the Danish system. This 4 set image has a target with a shirt image *. The numerical identification system 1 includes a captured ^ value of 4 7-knife image. The position device 15, a normalization device, a setting device 11, a conversion device 13, and a constant reading device 21 of the present invention. , A class of neural network 19, and a judgment In this embodiment, first, a specific image 3 is obtained by the user. The capture device 1 is used by the user. The capture device 11 can be any photographic device i and output to the conversion. Xiang home 13. This capture >

Yes, there are no restrictions here

As shown in Figure A of the second brother, ~ images and other images (the image with the target value π 3 displayed in Lingzhi image 31, is the instrument frame or other outside '7 ^ table is different). Here, other images may be like 31 or may be color or like. Because the captured specific image conversion device 13 will receive the last name 6 image, for further processing, the stage format is obtained after conversion, ie, temple conversion, device 15. Here, the above-mentioned two special features: image 33 'and send it to the positioning of any character representation image grayscale format refers to the 0 to plus device Further contains—storage disk (not: out). This storage device can be any memory,% memory, magnetic temporary storage 2 =, tape drive or other digital storage devices, for storage or I storage after conversion Specific image 3 3. As shown in the first figure B, the positioning device 15 of the present invention further includes a marginal device 1 5 1, a horizontal statistical device 1 5 3, a vertical statistical device, a horizontal positioning device 157, and a vertical Positioning device 159 'and a demarcation device 1 6 1. Here, because the contrast between the displayed characters of the specific image 3 3 and the moonscape is quite large,' Dang & bit device 15 receives the converted specific image After 33, it is convenient to mark all the edge parts of the converted specific image 3 3 with an edge device 1 5 1 to obtain a specific image 3 5 after being edged, as shown in the second figure c. At the same time, and separately Send it to the level

591545 V. Description of the invention (5) Leaf device 1 5 3 and vertical statistical vibration setting 1 5 5 As shown in the second figure C, after finding all the edge parts, a specific image 35 after S marginalization can be sent. Image of the edge part of the value = dense. Therefore, the horizontal counting device 153 and the vertical counting device 155 perform horizontal and vertical projections for the edge pixels of the edged image 35. As shown in the second figure D, that is, the cumulative statistics of the pixels in the horizontal and vertical directions are used to obtain a horizontal statistical result and a vertical statistical result. It can be seen from the second figure D that the target value part of the specific image 35 after marginalization will get a larger cumulative value. At this time, the statistical results are sent to the horizontal positioning device 157 and the vertical positioning device 59 respectively. Next, the horizontal positioning device 157 and the vertical positioning device 159 find out a horizontal area range and a vertical area range of the partial image containing the target value according to the horizontal and vertical statistical results, and send them to the delimiting device 161. Therefore, the demarcation device 161 can determine the size, range, and boundary of the partial image containing the target value in the specific image 35 after the edge region by analyzing the horizontal area range and the vertical area range to obtain a partial image containing the target value. 37, as shown in the second figure. In this embodiment, if the image 3 1 captured by the capture device 1 1 is over Λ, it is likely to get more than one large edge.疋 Η ^ ^ The area of the marginal pixels, when there are several possible partial image ranges with the target value, you can compare the aspect ratios of each range one by one to find the most likely. , This target value range can be used to filter other positioning ranges that are less likely.

Page 9 591545

V. Description of the invention (6), 隹 一 1: Human figure C / jinbu, ☆ In this embodiment, the normalization device 17 is a heading ::-a first device 171 and a second device 173. After finding out the range containing image 俨: 佶 Ϊ: divided image 37, the 'normalization device 17 receives the partial image 37 with the eye value, and the first device 171 normalizes the image 3? Two standard arrays with a fixed number of columns And output to the second device 173, the second device 173 normalizes the standard array from 0 to], sends it to a normalized image array, and outputs it to a neural-like network 9. Here, the above-mentioned fixed number of rows is set to 21 rows in advance. The number of rows of the standard array is determined according to the range-scale of the partial image 37 including the target value. But {, the size of the array set here is acceptable; it can be changed according to the needs of the user, and it is not used as a limitation. As shown in the first figure D, the neural network 9 of the present invention further includes an input device 191, an input layer 193, a first-scanning device 195, a first hidden layer 197, a second scanning device 199, _ The second 221, the second and third scanning devices 223, and an output layer 225. 'The input layer 193 has a first regional receptive field. The third hidden layer 197 has a second regional receptive field. The second hidden layer 221 has a third regional receptive field. In addition, here— In the embodiment, the mode excitation function of the neural-like network 19 used in the present invention is a sigmoid function. However, it is not limited here. When the neural-like network 19 is used by a normalization device! 7 receives this normal Image

Page 10 591545 V. Description of the invention (7) ^ ^ ^ Η 3 ^ 91 ^ "]# " In the same way, the first-hidden layer 197 follows its first! The same array is sequentially output to the second scan 2 field, and the reception is sequentially scanned and output to the first possession. Factory scanning device 199 according to its third regional receptive field;; Hidden layer 22 of array brother II] = m. The third scanning device outputs the layer 225. Therefore, by rotating your μ particles ’to the output layer 225, a set of analysis values is calculated. For example, as shown in the third figure + L, the part containing the target value in a specific image is identified as: 日 :: The normal two obtained by the required discrimination device 17 = Shed * 2 " 31) pixel outputs. In this case: =, the present invention divides the neural-like network 19 into four layers, which are an input layer 191], a middle fit layer 195, a second hidden layer 221, and an output layer 225. As shown in the third figure, the number of nodes from the input layer 191 to the output layer m is 231, 252, 140, and 10, respectively. Among them, the Tibetan layer 217 is divided into four groups. Each group has 63 nodes, and the groups of j are not connected. The second hidden layer 221 is also divided into four groups. 591545, invention description (8) Scholars However, since the neural network-like 19 system is quite large, in order to save computation = time, the present invention uses the so-called "local sexy = wild" (local receptive field) ), The complexity of the neural network 19. Here, in the present invention, the first layer of the input layer 19m is set, and the domain receptive field is 5 * 5, that is, the normalized pixel array is scanned with a 5 * 5 pixel array. However, due to the mortal and special nature of the value, the similarity of the pixels in the vertical direction of the target value will be superfluous. Therefore, in the present invention, scanning is performed in a manner of 2 pixels in the horizontal direction and 2 pixels in the vertical direction when scanning. Also like: two, the second regional receptive field of the hidden layer 197 is set to ㈣. The structure of the second hidden layer 221 and the output layer 225 is a setting that completely connects his regional receptive fields. As a result, as the other figure, as shown in the fourth figure, the input layer 193 and the first hidden layer ι97) are a 2 5 * 4 array. The __th class of neural network-like weights (an array of 1 4 0 * 1 0 between the W 2 reservoir layer 221 and the output layer 225. The network-like neural network between 9 and 9 The path weighting value (between the wj hidden layer 197 and the second hidden layer 22i) is a 36 * 4 array. The value of the second hidden neural network (W 3) is one point. The " of the output layer 225 of the present invention is 0, i, 2, 3, 4, ^, 8 and 9 which represent the Arabic numerals in sequence. After the output layer 225 performs calculation and analysis,

页 Page 12 591545 V. Description of the invention (9) 10 nodes obtain the -group analysis value 'and send it to the judgment device 21 As shown in the first figure E, the robust selection device 211 and -judgment device 213U1 of the present invention further include- The analysis value of the node, and select this group of analysis values $: f 21 1 Compare each 21 1 according to this maximum value, take the larger value against its location. Judgment device The number of eyes in this part of the image 37 containing the target value can be read out. It can be identified that the number of eyes in the specific image 31 is ^ 1. So, the present invention is done. 'Target numerical analysis of numerical partial images ^ 1 Bing Bing in the present invention' In addition to obtaining only one type of neural network with one type of neural network, it also uses several similar Randall's knives with different training samples to perform image recognition. And get the array analysis value. : Target; Ϊ: average the analysis results step by step, 'find the most accurate and objective fifth, and display the application of the numerical identification system of the present invention to identify one of a set of target values in a master image of a meter. example. The present invention can be applied to reading images of various meter devices 3, for example, any of a number meter, a water meter, a gas meter, and the like. There are no restrictions here. The numerical identification system 5 of this embodiment is different from the Emei system 1 of the first embodiment in that the numerical identification system 5 of this embodiment further includes a selection device 51 and a determination selection device 53.

Page 13 outside J :) V. Description of the invention (10) In this consistent embodiment, also from A Afc, one of the readings to be obtained is obtained, and the user uses the retrieval device 11 to retrieve the device U, which can be " ; 6]: Two outputs to the conversion device 13. This cut is' There are no restrictions on this. The child's clothes, or digital devices. But " 3 1 625, the scene = :: 1: 2 image 61 shows that the image with the -group S standard value may be an instrument frame or other, etc.) Here, 'its processing, conversion device " will J: early color image' & step-gradation format , T 刭, silk her: Jing Shu shirt like 61 is instantly converted into 25 6th order di one? The changed reading image 63 and transmitted to the fixed set 丨 5. Here, the above 256 order 255 The pixel of any character in the expression = means that it refers to the number from 0 to the device (other than the evening 1 '= send!) The numerical identification system 5 may further include a storage disc drive, light chan chan *, the storage device may be Any memory, register, magnetic temporary storage of the converted reading image 63. Display system, for storage or comparison ί two reasons:?:; = Display of number image 63 Character and background virtue, "Although the large bit device 15 receives the converted reading Mai Du 63, the edge part is equipped with f1: all / read the reading image 63, as shown in Figure C. When, And send them to the level separately

And vertical statistics device 155. As shown in Figure 6C, the mouth of all the edge parts is found in the readout image 65 after marginalization, and 1 of the target value can be sent.

The most dense. Therefore, the image of the horizontal statistics device 53 and the vertical and horizontal parts are respectively read from this marginalized image. The edge of the image 5 is exaggerated. The position is set to 1 5 5 ^ ^ «D direction is used to accumulate statistics to obtain a level statistical result. And — 1 party. As can be seen from the second figure D, the marginalized readings will result in a larger cumulative value for the 5-knot value. It is expected that the horizontal positioning device 157 and the vertical positioning device 159 will be used. =: The device 157 and the vertical positioning device 159 are based on the horizontal and vertical two: Find the I target value part of the image-the horizontal area range is ^ a vertical area range, and send it to the demarcation device 61. Therefore, the boundary analysis of this horizontal area range and this vertical area range, and the bounds and boundaries of the target value part image in the image 65, to obtain a target value part image 67, as in the sixth embodiment in this embodiment, If the image 6 1 is too large, the area will likely be obtained, that is, there are several possible time-lapses. You can compare the range of the scalar value of each range one by one, and you can filter more than the readings obtained by the lifting device 11 The target value of the larger edge pixel of the partial image range. This aspect ratio finds the most likely and less likely positioning range.

Page 15 591545 V. Description of the invention (12) Next, this embodiment further uses a selection device 51 to sequentially select a specific image 671 of a fixed size from the partial image 67 containing the target value, and inputs the specific image 671 to the normalization device 17. In this embodiment, the specific image 671 in the partial image 67 including the target value is selected from left to right. When the selected specific image 671 has completely covered any-specific target value, the 分析 i of the 10 analysis values of the layer m of the round of the neural network i 9 in the first I embodiment and the like is set to a large standard. difference. At this time, the selection device 51 sets this to the normalized dress 17. The specific image 671 refers to the target-specific value among the target values including the binary content. Therefore, the second one: "The size of the temple image 671 is fixed approximately to this specific target number. Buddle ΓΓ; / In the embodiment, this normalization device 17 further includes-the first clothing device 171 and a second device 173. . Zhengdi ~ have a fixed 歹 I #huai hit into at least: two devices 173 to this standard array _ normal two 3. In the example, and output to the neural network 19. Here-* purely according to the target The numerical part of the image 67 is determined by the :: rate matrix dagger t. However, the array set here is subject to change and is not used as a limitation. Page 16 V. Explanation of the invention ( 13) The type excitement: number: column 'The analogous neural network 19 used in the invention: the analogous neuron: 丨 1 two numbers. However, after this is not used as a limitation, the input event is set to 191. ^ / 17 # Receive this normalized image array into the layer, input; like array, sequential out to the first-scan: set 1 first to receive, A clothes will be scanned in order, and output to the first barrier reed q 7 First, the first hidden layer 197 is the 1st m: subtracted layer 197. The same array is sequentially rotated out until the second scan is set to V = Q, and the sequential scan will be received. The output is to the second hidden layer 221. ^ ―The scanning device m then its third regional experience #, and the receiving layer 221 is set according to the scanning device 223. The third scanning device 2 sequentially outputs to the third scanning layer. Therefore, the output layer 22 5 will output two f to output 225, and then send this set of analysis values to the interpretation I]. The output layer 211 compares 1 and 3 to this set of analysis values, which is convenient. Select the analysis value of the device letter Λ and select the largest judgment device in the set of analysis values. 2 1 1 Based on this maximum value, set It to the target value of It. At this time, find the specific value in the shirt image 6 71 The target value. The present invention further includes using a determination and selection device 53 to determine whether all the images 67 with a value of 4 points have been completely selected by the selection device 51. 5. The description of the invention (14) is completed. Set 51 to continue to select another-specific image, so that 4 f3 will be notified to select the value. Until the target value Qiu Ba and the next f identification 4 another specific target number is known, this set of target values are all; = finished, also In addition to the value to be analyzed, a can be: two T is only based on-like neural network Get-group neural network, enter J two numbers separately: J has different training samples: after, then-step further to :::: flat; analysis value. The target value of. Use. If the readings sent by the transmission device can achieve this, you can do it if you do n’t get it. For the company can make this profit. In addition, the seventh picture is 71. A remote image 6 3 The invention is for anyone who wants to know the power It is convenient for some companies. 'The present invention also shows that the transmission limit of the function of the data processing of the interpretation device 21 is directly transmitted. Therefore, the target value is provided for the company, _ means that the application can be easily further Ming further included the identified position 73. In addition, the remote data can be sent by this method. Here, the tap water provides the transmission-contained device 71 for obtaining a long distance as a long-distance identification value mentioned in the present invention. In addition to this specific target value, the post-conversion processing position 73 can also be used. And various are applicable in this system. You can easily use this data processing position 7 3 'This data processing position 7 3 company, or a gas identification system provided by the customer, is the consumption value of the client, according to 591545 V. Description of the invention (15) In order to explain the detailed steps of the numerical identification method of the concept of the present invention more clearly. . 2 I · f i, a first implementation of the numerical identification method of the present invention. Two goals: Kit: Kit: Department: To identify the numerical value of the invention in a specific image, and mark the numerical partial image. The method for distinguishing the age includes steps 801 to 815. Digital image: = system; capture a specific image. In this embodiment, this reading, or 9 2 = how to capture, for example, using a photographing device, an optical device, an image package containing the letter r, is not a limitation. And this particular can be several images of the instrument and other images. Other images can be busy or other external images. Or κ Α because the specific image captured in step 803 may be in color. For further processing, in step 8 05, the 4 after receiving is fixed / converted to 25 6-level gray scale format. At the time of this writing, the 25-level gray scale format refers to the pixel format that is represented by any sub-70 in the range 0 to 1. In addition, this method can store this specific image for further confirmation. Step 80 ”is here, step 8 ° 7-step by step into step 8081. Baixian, step 8 () 71 is after receiving this conversion

Part I on page 19: After the image is fixed, locate the target image with the value 591545 5. After the specific image of the invention description (16), the converted image will be shown. Because the converted specific image 2 ~ all the edges are marked quite large '因 &' One side can be obtained '::: 70 After the contrast with the background has an edge part, ⑽; after finding the edge of the specific image after the fate; for = cast: step: 73 is added to this side The pixel μ of this edge part also finds a result; it is the pixel that accumulates the edge part horizontally and vertically ^ = step 8075, the target value part of a specific image will be = fruit. Scaled out from the range: == =; large pixel cumulative image size, r heart shadow:; = part of the statistical results of the scalar value, find out the person's day slave / ± person ν8080 is based on this horizontal step, the system According to: vertical statistics =: ”=: wide ° 81 is obtained according to the above:“ The range of the part of the image that contains the target value in the specific image. The sacrifice sacrifice will be very practical. Because the specific image captured is too large, there are several = 2 ^ to more than one larger edge pixel area, that is, + hand girl β = part of the image range with the target value. At this time, you can go through the aspect ratios one by one: find the most likely target value range, and filter out other more unlikely positioning ranges. Page 20 591545 V. Description of the invention (17) ^ Next, step 80 9 is to find out the range of the part image containing the target value and convert the part image containing the target value into a normalized image array. As shown in the tenth figure, step 809 is further divided into steps 8091 to 8093. Step 80 9 1 normalizes a specific image into at least one labeled array with a fixed number of columns. Steps 80 and 93 normalize the standard array to obtain a normalized = image array. In this embodiment, the above-mentioned fixed number of rows is predetermined. The number of rows in this standard array is determined according to the size of the part containing the scalar value, and is determined according to a certain ratio. However, it is limited here: the size of the array can still be changed according to the needs of the user, and this is not used. ^ # TAfter receiving this normalized image array, step 811 inputs &# π: the image array to a class of nerves On the Internet, I have analyzed the value of this regular set of daggers. In this embodiment, the second embodiment of the present invention uses the neuron-like system of 4 si x jin as the excitation function of the model of the sub-fu to road. gm〇ld function. However, it is not limited to this. In addition, the neural network of the present invention includes a hidden layer, a second hidden layer, a turn-in layer, and a first contact regional receptive field. The first-hidden layer has two or two: the input layer has a first one, the Tibetan layer has a third region, and a regional receptive field. Step 811 is further divided into steps; As shown in Fig. 10—all *, in this order, this normalized image array is "in accordance with the _ system according to the -layer j = class; f round of layers. Step 丨 Wang Zhuoye wild, scan wheel 591545 V. Description of the invention (18) Transform the image array, and turn to the first hidden regional receptive field, scan input to the Zang: Yi step 8115 is based on the second brother to the second Hidden layer. Step 8U7 is an image array of the day-by-day layer. Solitary: Draw the image input to the second hidden layer ~ the three regional receptive fields, and calculate a set of analysis values. 4 'and it is rounded out to the output layer, and the outer step 813 is based on the obtained target value. As in the twelfth, the analysis value is judged to read the desired step 81 31 to step 81 33. Step 81 = j Step 813 is further divided into steps to compare these analysis values, and select the analysis value obtained from this set of analysis values, and then compare the maximum value with the maximum value according to this maximum value. Step 8 1 33 The target value in this specific image, 1 and 1 Arabic number, 'Judge the target of the part of the image containing the target value in the read image to identify the specific. In addition, in the present invention, in addition to only one analysis value, You can also get a set of neural networks at the same time, and perform image recognition separately = class of training samples; after that, further analyze the results by averaging :: analysis values. Target value. Accurate and objective. In addition, the thirteenth figure shows the application of the present invention: the identification of a group of target numbers in a reading image in a meter device: an example. The present invention can be applied to the identification of various meter devices, such as readings, electricity meter, water meter, and gas meter. : This, = page 22 591545 V. Description of the invention (19) There are no restrictions. The numerical identification method of the present invention includes steps 丨 3 〇 to steps 1 3 1 9. The difference from the first embodiment is that the numerical identification method of this embodiment further includes steps 1309 and 1317. Baixian v 1. Take a few pictures of you. In this embodiment, the reading image can be captured in any manner, such as using a device, an optical device, or a digital device. And this reading image shows the image of 乂 ^ value and other images. Other images may be outside the instrument 2: Other external images. The frame of the smart meter is either ^ and 'because of the reading scene captured in step 3 03 or a monochrome image, for further processing, step 2 = may be a grayscale format of 2 56 levels, Obtained—the converted reading 05 is converted immediately to the above-mentioned grayscale format of 256, which means that 0 to 255. Here, the pixel format is shown. In addition, this method can further character image for further confirmation. Giving this reading image step 1 3 0 7 is to divide the image based on the converted reading image. This step is more detailed than the first embodiment. The method contains the same number of targets. The difference between this example and the first Shibaishi example is that, here, Yumi = the target value determined from step 1 3 07; Specific image. In this implementation image, 丨

Page 23 591545 V. Description of the invention (20) The specific image in the numerical part of the image is selected from left to right. When the selected specific image completely covers any specific target value, the set of analysis values obtained by the output layer of the neural network like the above will obtain the maximum standard deviation. The specific image refers to a fixed-size image among the images containing the target value. A specific image contains a specific target value from this set of target values. Therefore, the present invention sets the size of the specific image to be fixed to approximately the size of the specific target value. Step 1 3 1 1 then converts the part of the image containing the target value into a normalized image array. Step 1 3 1 3 Enter this normalized image array into a type of neural network for computational analysis and obtain a set of analysis values. Step 1 3 1 5 Based on the obtained set of analysis values, judge and read out the specific target value you want to know. The above steps 1 3 11 to step 1 3 1 5 are the same as those in steps 809 to 8 1 3 in the previous embodiment. Here, no more details are provided. After step 1 3 1 5, the method further performs step 1 3 1 7. Step 1 3 1 7 is to determine whether this part of the image with the target value has been completely selected in step 1 3 0 9. If the selection is not complete yet, the method proceeds to step 1309 to continue selecting another specific image in order to identify another specific target value. Until the part of the image with the target value is selected, also

Page 24 591545 V. Description of the invention (21) It is to know that the target value of this group has been identified. In addition, this method can be further applied to long-distance identification. Fig. 14 shows a third embodiment of a long-distance numerical identification method of the present invention. The present invention includes steps 1 401 to 1 421. Step 14 1 to step 1 4 1 5 are the same as steps 1 3 0 1 to 1 3 1 5 in the second embodiment. The difference from the second embodiment is that 'this embodiment further includes step 1417. This step 1 4 1 7 further transmits the specific target value identified in step 1 4 1 5 to a remote data processing location. In addition, the converted reading image can also be transmitted directly to the remote data processing location. And various transmission methods that can reach the functions of the invention can be applied to this system. There are no restrictions here. Therefore, the target value data can be easily obtained from the data processing position. Here, this data processing bit ^ ^ power mention, company,-water supply company, or 1 siti = 1. In other words, by applying the numerical identification system provided by the present invention, some companies can easily obtain the consumption of the remote user terminal, which is convenient. Feng Yili st .. Ke Mi Chengming's book, the specific embodiment Aj of this creation is also not described, but obviously various modifications and changes do not depart from the spirit and scope of the second. The corresponding descriptions and illustrations are used to limit the scope of this creation. Therefore, I 17 starts with the description and the Dagger 7 " This creation should cover all occurrences 4

Page 591 545

Page 26

591545 Brief description of the drawings The first diagram A shows a block diagram of a first embodiment of the numerical identification system of the present invention; the first diagram B shows a block diagram of a positioning device of the first embodiment of the numerical identification system of the present invention; A figure C shows a block diagram of a normalization device of a first embodiment of the numerical identification system of the present invention; a first figure D shows a block diagram of a neural network-like network of a first embodiment of the numerical identification system of the present invention; The first figure E shows a block diagram of a reading device of a first embodiment of the numerical identification system of the present invention; the second figure A shows a schematic diagram of a specific image captured by a first embodiment of the numerical identification system of the present invention; The second figure B is a schematic diagram of a converted specific image of a first embodiment of the numerical identification system of the present invention; the second diagram C is a schematic diagram of the specific image after marginalization of a first embodiment of the numerical identification system of the present invention; The second figure D shows a schematic diagram of positioning a part of the image containing the target value in a first embodiment of the numerical recognition system of the present invention; the second figure E shows the numerical recognition system of the present invention A schematic diagram of a part of a target embodiment including a numerical value of the first embodiment; FIG. 3 shows a schematic diagram of a neural network-like structure of a first embodiment of the numerical identification system of the present invention; and FIG. 4 shows a schematic of a numerical identification system of the present invention. Schematic diagram of the neural network structure of the first embodiment;

Page 591545 Brief description of the diagram The fifth diagram shows a block diagram of a second embodiment of the numerical identification system of the present invention; the sixth diagram A shows the readings taken by a second embodiment of the numerical identification system of the present invention Image schematic diagram; FIG. 6B shows a converted reading image diagram of a second embodiment of the numerical identification system of the present invention; FIG. 6C shows a marginalized image of a second embodiment of the numerical identification system of the present invention. Schematic diagram of the reading image; FIG. 6D shows a schematic diagram of positioning a part of the image containing the target numerical value in a second embodiment of the numerical identification system of the present invention; and FIG. 6E shows a schematic diagram of a second embodiment of the numerical identification system of the present invention. Schematic diagram of the target numerical partial image and specific image; FIG. 7 shows a block diagram of a third embodiment of the numerical identification system of the present invention, and FIG. 8 shows a flowchart of a first embodiment of the numerical identification method of the present invention; The figure shows a flow chart of a conversion step in a first embodiment of the numerical identification method applying the present invention; the tenth figure shows the numerical values applying the present invention A flowchart of a positioning step of a first embodiment of the identification method; FIG. 11 shows a flowchart of a normalization step of a first embodiment of the numerical identification method of the present invention; FIG. A flowchart of a kind of neural network analysis steps of a first embodiment in the numerical identification method of the invention,

Page 591545 Brief Description of the Drawings Figure 13 shows a flow chart of the second embodiment for identifying a set of target values in a reading image in a meter device using the numerical identification method of the present invention; and Figure 14 shows A flowchart of a third embodiment of a method for long-distance numerical identification according to the present invention. Explanation of Symbols of Graphical Elements 1 Numerical identification system 151 Marginalization device 3 Instrumentation device 153 Horizontal statistical device 5 Numerical identification system 155 Vertical statistical device 11 Acquisition device 157 Horizontal positioning device 13 Conversion device 159 Vertical positioning device 15 Positioning device 161 Boundary device 17 Normalization device 171 First device 19 Type neural network 173 Second device 21 Interpretation device 191 Input device 51 Selection device 193 Input layer 53 Confirmation selection device 195 First scanning device 71 Transmission device 197 First hidden layer 73 Data processing location 199 Second scanning device 31 Specific image 221 Second hidden layer 33 Transformed specific image 223 Second scanning device 35 Marginalized specific image 225 Output layer 37 Partial image with g-scalar value 211 Selection device

Page 29

591545

Page 30

Claims (1)

591545 VI. Scope of Patent Application '1. A numerical identification method using a neural network is used to identify a target value in a specific image, the specific image includes a partial image containing the target value, and the numerical identification method includes the following Steps: (a) capture the specific image; (b) convert the specific image into a grayscale format; (c) locate the partial image containing the target value according to the converted specific image; (d) convert the target value Part of the image is a normalized image array; (e) input the normalized image array to a type of neural network for computational analysis and obtain a set of analysis values; and (f) judge and read out the target value based on the set of analysis values . Bu Hai = the number of discernment values in the mouth of the law: The first step of the first step is 乂 Θ, the “changer” subcontracted, and the special step is stored, such as (b) u 2 (b) The special post fanli specifically requested Shen Ru (C 3. The progress of the legal discernment value in the step: the step 1 under step 1 includes the step and the image included; the shadow fruit points •, the end part The evaluation unit counts a standard edge to the image that is included in the target, and the image is fixed and fixed. The branch of the fruit is converted to the trailing edge, and then the data is found to find the root. The number of legal discernment values mentioned in the step is 3 Page 31 591545 6. The scope of patent application (c.2) further includes the following steps: (c. 2 1) horizontally accumulate the pixels of the edge part to obtain a horizontal statistical result; and (c. 22) vertically accumulate the edge Part of the pixels, a vertical statistical result is obtained. 5. The numerical identification method as described in item 4 of the scope of patent application, wherein step (c.3) further includes the following steps: (c. 31) According to the statistical results of the level, find a A horizontal area range; and (c. 32) based on the vertical statistical result, find a vertical area range of the partial image containing the target value. 6. The numerical identification method as described in item 5 of the scope of patent application, wherein step (c.3) further includes the following steps: (c. 3 3) Analyze the horizontal area range and the vertical area range 'to determine the inclusion The range of the partial image of the target value. 7. The numerical identification method as described in item 1 of the patent application scope, wherein step (d) further includes the following steps: (cL 1) normalizing the partial image containing the target numerical value to at least one standard array with a fixed number of columns; and ((L 2) normalizes the standard array into a normalized image array. Page 591545 Among them, in the scope of patent application 8. The numerical identification method step (d · 1) described in item 7 of the scope of patent application, the fixed number of columns is 21 columns. Wherein the numerical value identification method according to item 7 of the scope of the patent application, in step (dl), the number of rows of the standard array is based on the range of the image with the subdivision and according to a predetermined ratio Determined 10 · If the first neural network of the scope of the patent application includes an input layer, an output layer, and a turn-in layer (receptive field), the first field of the $ 's second hidden layer has a further step including: : (E · 1) sequentially input the normalized (e · 2) according to the first regional image array, and output to the (e · 3) image array according to the second regional layer, and output to ( e · 4) according to the image array of the third regional layer, and output to the numerical identification method, wherein a first hidden layer, a second hidden layer of this type has a first regional receptive field (1 〇cal A hidden layer has a second regional receptive field and a second regional receptive field. In step (e), the image array is input to the input layer; the receptive field is scanned into a hidden layer of the input layer; the receptive field is scanned into the input. To Cain a hidden layer; and a receptive field, scan the input to the second hidden output layer to obtain the set of analysis values. 11 · The numerical identification method according to item 1 of the patent application scope, wherein the step (f ) Further includes the following steps: (f · D compares the analysis values of the group, and selects one of the maximum values of the analysis value of the group; Page 591 545 1 2 · According to the numerical identification method described in item 1 g ^ ^ ^ 1 of the scope of the patent application, I. This type of neural network further describes Dong ^, where the neural subnet, step (e) proceeds-step Contains: The sample (e · 5) of the sample book uses the τ < Gongbei neural network to obtain the array analysis values respectively; and (e · 6) selects an appropriate value from the analysis values. A set of analysis values. 1 3 · —A numerical identification method using a neural network is used to identify a set of target values in an image. The image includes a partial image containing the target value. The numerical identification method includes the following steps: (a) operation Take the image; (b) convert the image into a grayscale format; (c) locate the partial image containing the target value according to the converted image; (d) select a specific image among the partial image containing the target value, the The specific image contains a specific target value in the set of target values; (e) converting the specific image into a normalized image array; (f) inputting the normalized image array into a type of neural network for computational analysis and obtaining a set of analysis values; and (g) according to the set of analysis values, Judging the specific target value. 1 4 · The numerical identification method according to item 丨 3 of the patent application scope, wherein Page 34 591545 VI. The method for identifying the range of patent application values further includes the following steps: (h) determining whether the part of the image containing the target value in step (d) has been completely selected; and (i) when the target value is included This step (d) is performed when some images have not been completely selected. 15. The numerical identification method as described in item 13 of the scope of patent application, wherein the specific image refers to a fixed-size image of the partial image containing the target numerical value. 16. The numerical identification method as described in item 13 of the scope of patent application, wherein step (b) further includes the following steps: (b · 1) storing the converted image. 1 7. The numerical identification method described in item 13 of the scope of patent application, wherein step (c) further includes the following steps: (c · 1) finding an edge portion of the image after conversion; (c. 2 ) Counting the pixels of the edge part, and get a statistical result; and (c. 3) According to the statistical result, locate the range of the image containing the target numerical part. 1 8. The numerical identification method as described in item 17 of the scope of patent application, wherein step (c.2) further includes the following steps: (c. 2 1) Horizontally accumulate the pixels of the edge part to obtain a horizontal statistical result. Page 35 591545 VI. Results of patent application; and (c. 22) Vertically accumulate pixels of the edge part to obtain a vertical statistical result. 19. The numerical identification method as described in item 18 of the scope of patent application, wherein step (c.3) further includes the following steps: (c. 3 1) According to the statistical result of the level, 'find the part with the target value A horizontal area range of the image; and (c. 3 2) finding a vertical area range of the partial image containing the target value according to the vertical statistical result. 20. The numerical identification method as described in item 19 of the scope of patent application, wherein step (c.3) further includes the following steps: (c. 3 3) Analyze the horizontal area range and the vertical area range ^ to determine The range of the partial image with the target value. 2 1. The numerical identification method according to item 13 of the scope of patent application, wherein step (e) further comprises the following steps: (e. 1) normalizing the specific image into at least one standard array with a fixed number of columns; and (e. 2) Normalize the standard array to a normalized image array. 2 2. The numerical identification method described in item 21 of the scope of patent application, wherein in this step (e · 1), the fixed number of columns is 21 columns. Page 36 591545 VI. Patent Application Range 2 3. The numerical identification method described in item 21 of the patent application range, wherein in this step (e. 1), the number of rows of the standard array is based on the target value. The size of the partial image is determined according to a predetermined ratio. 2 4. The numerical identification method according to item 13 of the scope of the patent application, wherein the neural network includes an input layer, a first hidden layer, a second hidden layer, and an output layer. The input layer has A first regional receptive field, the first hidden layer has a second regional receptive field, and the second hidden layer has a third regional receptive field. This step (f) further includes the following steps: (f · 1 ) Sequentially input the normalized image array to the input layer; (f · 2) scan the image array input to the input layer and output to the first hidden layer according to the first regional receptive field; (f · 3 ) Scan the image array input to the first hidden layer according to the second regional receptive field and output to the second hidden layer; and (f. 4) Scan and input to the third regional receptive field according to the third regional receptive field. The image array of the second hidden layer is output to the output layer to obtain the set of analysis values. 2 5. The numerical identification method as described in item 13 of the scope of patent application, wherein step (g) further includes the following steps: (g. 1) compare the analysis values of the group and select one of the maximum values of the analysis value of the group ; And (g. 2) judge and read out the specific target value based on the maximum value. Page 37 591545 VI. Patent application scope 2 6. The numerical identification method described in item 13 of the patent application scope, wherein the neural network further includes several neural-like subnetworks each having different training samples, the Step (f) further includes: (f. 5) using the plurality of neural-like sub-networks to obtain array analysis values respectively; and (f. 6) selecting an appropriate set of analysis values from the plurality of analysis values . 2 7. The numerical identification method according to item 13 of the scope of patent application, wherein in step (a), the image is obtained by a photographing device. 2 8. The numerical identification method described in item 13 of the scope of patent application, wherein in the step (a), the image is obtained by an optical device. 2 9. The numerical identification method described in item 13 of the scope of patent application, wherein step (a) further includes the following steps: (a. 1) transmitting the image to a data processing device by a predetermined method. 30. The numerical identification method described in item 13 of the scope of patent application, wherein step (g) further includes the following steps (g. 3) transmitting the specific target value to a data processing by a predetermined method Device. 3 1. —A method of identifying instrument readings using a neural network Page 38 591545 VI. Scope of patent application Identifying a reading image on a meter device The image contains a target value and part of the image contains the following steps: Two sets of target values, the reading, the meter reading identification method package (a) (b) (c) (d) ( e) capture; capture the reading image; convert the reading image into a figure that finds the converted reading and count the edge portion of the image; determine the grayscale format according to the statistical result; an edge portion of the image; and get a statistic Results: The range (f) of the partial image with the target value is selected. The partial image with the target value is included in the set of target values. (G) The specific image is normalized as a column. A specific target value of a specific image in the partial image ; At least one with a fixed number of columns' this particular standard array (h) normalizes the standard array as-normalized image array for calculation (i) inputs the normalized image array to a class of neural network analysis and obtains a set Analysis value; (j) comparing and selecting one of the maximum values of the group of analysis values; (k) judging and reading a specific target value according to the maximum value, (l) judging in step (f) The partial image with the target value is completely selected; and & white dagger (m) when the partial image with the target value has not been completely selected, step (f). 〗 仃 3 2 · The meter reading identification method described in item 31 of the scope of patent application, which Page 39 : == Fixed image means that one of the partial images containing the target value is fixed at 3 3 7 • The meter reading identification method described in item 31 of the patent application scope, step (b) in 1 further includes the following steps: "(B · 1) Store the converted image of the reading. = · This step (d) in the meter reading identification described in item 3 丨 of the patent application scope further includes the following steps: (:, 1) Accumulate the pixels of the edge part to get a horizontal statistical knot (d · 2). Accumulate the pixels of the edge part vertically to get a result. Accurate vertical statistical knot e) further includes the following steps: a number identification method, e · I) a horizontal area range according to the horizontal statistical result; and (e · 2) a vertical area range according to the vertical statistical result. Finding the target The numerical partial image finds the partial numerical image with the target value. 36. As described in Item 35 of the scope of the patent application ~ This step (e) further includes the following steps: (e · 3) Analyze the horizontal area range and the vertical area Range to determine Page 40 591545 VI. Patent Application Range This is the range of part of the image containing the target value. 37. The meter reading identification method as described in item 31 of the scope of patent application, wherein in this step (g), the fixed number of rows is 21 rows. 3 8. The meter reading identification method according to item 31 of the scope of patent application, wherein in step (g), the number of rows of the standard array is based on the range of the partial image containing the target value, and according to a predetermined ratio Decided. 39. The meter reading recognition method according to item 31 of the scope of patent application, wherein the type of neural network includes an input layer, a first hidden layer, a second hidden layer, and an output layer, the input layer There is a first regional receptive field, the first hidden layer has a second regional receptive field, and the second hidden layer has a third regional receptive field. The step (i) further includes the following steps: (i · 1) sequentially input the normalized image array to the input layer; (i. 2) according to the first regional receptive field, scan the image array input to the input layer and output to the first hidden layer; (i · 3) Scan the image array input to the first hidden layer according to the second regional receptive field and output it to the second hidden layer; and (i. 4) Scan and input to the third regional receptive field according to the third regional receptive field. The image array of the second hidden layer is output to the output layer to obtain the set of analysis values. 40. The meter reading identification method described in item 31 of the scope of patent application, which Page 41 591545 6. In the scope of the patent application, this type of neural network further includes several neural-like sub-networks with different training samples. This step (i) further includes: (i. 5) using the several neural-like networks. Subnets to obtain array analysis values; and (i. 6) selecting an appropriate set of analysis values from the array analysis values. 4 1. The meter reading identification method described in item 31 of the scope of patent application, wherein step (a) further includes the following steps ... (a.1) By a specific method, the reading image is transmitted to a data processing device. 42. The method for identifying meter readings according to item 31 of the scope of patent application, wherein step (k) further includes the following steps: (k. 1) transmitting the specific target value to a data processing device by a specific method. 4 3. —A method for identifying remote meter readings using a neural network, which is used to identify a set of target values in a reading image on a meter device and send it to a remote data processing device. The reading image The method includes the following steps: (a) capturing the reading image; (b) transmitting the reading image to the data processing device by a specific method; (c) Convert the reading image into a grayscale format; Page 42 591545 6. Scope of patent application (d) Find the conversion (e) Horizontal accumulation (f) Vertical accumulation (g) According to the water-horizontal area norm (h) According to the vertical-vertical area norm (i) The target (j) selects the image package schedule; (l) the positive (m) input analysis, (η) is better than (0) the root (P) is judged to be complete (q) The group standardizes the regularization, the normalization, and the comparison and selection based on the most discontinued step selection; the one with the edge (j) 〇 = of the Xuanjing number scene > an edge portion of the image; Pixels, get-the image of the horizontal part of the horizontal edge *, get-weeping: ten, and fruit; ten statistical results, find out the number of images containing the objective I ,, and fruit; know the numerical part of the image Straighten the statistical results to find the range containing the target; know the flat range of the numerical partial image and the range of the vertical sub-image;-Siege to determine a specific target in one of the target values in the target numerical partial image; :: Image, the specific special image is at least one standard array with a fixed number of columns. The standard array is a normalized image. The image array normalizes the image array to a set of analysis values of a class of neural networks; choose one of the maximum of the set of analysis values; a large value 'judges to read out the specific target value; whether the part of the image containing the target value in step (j) is And if the partial image of the target value has not been fully selected, then proceed to calculate J 44. The distance meter reading identification method described in item 43 of the scope of patent application P.43 591545 VI. Method (c 45 Method No. 1) (m (m 影 (m layer (m layer patent application scope-, where step (c) further includes w τ Zhi • 1) After storing the conversion The reading image of the reading. • As described in item 43 of the scope of the patent application, the neural network of this type includes a hidden layer and an output layer, a first hidden layer, and a field, The first-hidden layer has a layer-the first regional reservoir layer has a third regional receptive field— & regional receptive field, the second next step: 'This step (m) further includes • 1) ^ order Input the normalized image array to the input sound. • 2) According to the first region, the image input is input to the input column of the input layer and output to the first hidden layer; ^ ^ According to the second regional receptive field, scan input to the first hidden shirt image array and output to the second hidden layer; and 4 L according to the third regional receptive field, scan input to the second hidden field Shirts are like an array and output to this output layer to To the group of analysis values. 4 6 · As described in the patent application scope * 3 households described the remote meter reading identification method 'where this type of neural network includes / less includes several neural-like subnets with different training samples (M) further includes: (m · 5) using the plurality of neural-like subnetworks to obtain array analysis values respectively; and (16) selecting an appropriate set of analysis from the array analysis values value. 47. According to the application of the long-distance meter reading in item 43 of the application, the far-distance special method is to use a green, phase m +, Ί,…, line network system through at least one resource. (Data channel) transmission. Wang / Bei 48 ·-In order to distinguish the far part (a) (b) (c) (d) (e) (f) a water (g) a vertical ~ a kind of data using a neural network-like telemeter end The processing device sends the image and the target value to the image, and the remote meter reading discriminator J second standard value captures the image of the reading; D The following steps: • Convert the far-term number image into a grayscale format • Find out the conversion After this reading, Shuifu knows an edge part of the 1Γ image; water accumulates the image of the edge part and vertically accumulates the horizontal statistical result of the edge part; according to the horizontal statistical result, find 哕:, two vertical statistical results; Flat area range; ^ S target value part of the image based on the vertical statistical results, straight area range; 3 mesh 'value part of the image (h) analysis of the horizontal area range to determine the' this particular standard matrix contains the target value part of the image The range of the straight-line range β (0. Select the image with the target value to contain the set of target numbers. Φ knife ~ one of the image-a specific scene to normalize the specific image to at least one value, column; ~ King / one with a fixed frame, (k) Regular The standard array Chihaya bad Shu is a normalized image array; 591545 6. Scope of patent application (l) Input the normalized image array to a type of neural network for calculation and analysis, and obtain a set of analysis values; (m) Compare and select one of the maximum of the set of analysis values; ( η) judge and read the specific target value according to the maximum value; (0) send the specific target value to the data processing device by a specific method; (P) determine the part with target value in step (i) Whether the images have been completely selected; and (q) when the partial image with the target value has not been completely selected, then perform step (i). 4 9. The remote meter reading identification method as described in item 48 of the scope of patent application, wherein step (b) further includes the following steps: (b. 1) storing the converted image of the reading. 50. The remote meter reading recognition method according to item 48 of the scope of the patent application, wherein the neural network includes an input layer, a first hidden layer, a second hidden layer, and an output layer. The input The layer has a first regional receptive field, the first hidden layer has a second regional receptive field, and the second hidden layer has a third regional receptive field. The step (1) further includes the following steps: (1 1) Sequentially input the normalized image array to the input layer; (1. 2) Scan the image array input to the input layer according to the first regional receptive field and output to the first hidden layer; P.46 591545 VI. Applying for scrolling ---- ^ Xi Shaoyi knows that the second area is sexy and wild, scans into the shirt image array of the first hidden layer, and turns out to the second hidden layer; And (1 · 4) scanning the image array input to the second hidden layer according to the third regional receptive field, and rotating out to the output layer to obtain the set of analysis values. 51. The method for identifying remote meter readings as described in item 48 of the scope of the patent application, wherein the neural network includes at least several neural-like subnetworks with different training samples, and (1) further Contains: (1 · 1) using the neural neural network to obtain array analysis 傧. And (1.2 3 4) from the analysis value of the array, select an appropriate set of analysis values. 5 2 · The remote meter reading identification method as described in item 48 of the scope of patent application, wherein the specific method is utilized, and the wireless network system transmits through at least a data channel. A positioning device for positioning the _ target value part of the image based on the specific image after conversion; ^ 含 目 1 3. —A numerical identification system using a neural network is used to identify a target value in a specific image. The specific image includes a 5 part image with a visual instruction. The numerical identification system includes: A capture device for capturing the specific image; 5 a conversion device for converting the specific image into a gray scale pattern; 591545 VI. Patent application image array; a type of neural network, the normalized image array is input to the class A neural network to perform computational analysis and obtain a set of analysis values; and a reading device for judging and reading the target value based on the set of analysis values. 5 4. The numerical identification system according to item 53 of the scope of patent application, wherein the numerical identification system further comprises: a storage device for storing the converted specific image. 5 5. The numerical identification system according to item 53 of the scope of patent application, wherein the positioning device further comprises: an edge device for finding an edge portion of the specific image after conversion; a horizontal statistics device for Horizontally accumulate the pixels of the edge portion to obtain a horizontal statistical result, and a vertical statistical device for vertically accumulating the pixels of the edge portion to obtain a vertical statistical result. A horizontal positioning device, based on the horizontal statistical result, finds the target containing the target. A horizontal area range of the numerical partial image; a vertical positioning device that finds a vertical area range of the target numerical partial image based on the vertical statistical result; and a demarcation device that analyzes the horizontal area range and the vertical area range, To determine the range of the part of the image with the target value. Page 48 591545 VI. Application for patent scope 56. The normalization device described in item 53 of the scope of patent application further includes: identification system, where: :: 士士 'is used to normalize the target number with a fixed number of columns A standard array; and ~ is a at least one second device for normalizing the standard array to a normalized image array. 57. The number of fixed columns is 21 as described in item 56 of the scope of patent application. Identification system, where the number of 5 8 · as described in Item 56 of the scope of the patent application ... The number of rows of the standard array is based on the identification system with the item 'where it is small, and determined by a predetermined ratio T ° 卩 image The range is large. 59. The neural network described in item 53 of the scope of patent application includes an input layer and a recognition system, wherein the Leyi hidden exhibition, A # layer, and an output layer, the input layer has a 9 A second concealment The first concealed layer has a second region & Sense II ^ a regional receptive field, there is-a third regional receptive field, this type of nerve :: enter =: hidden layer with-input equipment, supply Sequentially input the image array of the normal input layer, and output to the first hidden ^ Putian round to-the second scanning device 'according to the second regional receptive field: Scan Hi: ΐ image array', and output To the second hidden layer "is set with i-the second ... according to the third regional receptive field, scan lose = 苐 Page 49,591545 VI. The analysis range of the patent application group is the image array of the second hidden layer, and it is output to the round-out layer to obtain the 6 0. If the patent application scope of the 53rd interpretation device further includes: Numerical identification system, in which the selection device, # 屮 屮 西 &, 丄 丄 乂 乂 分析, the set of analysis values, to select #, one to take a larger value; and ^ to choose one of the set of analysis values-judgment device ' Based on this maximum value, the J target β value is obtained. 61. The numerical neural network according to item 53 of the scope of patent application further comprises a plurality of = systems, wherein the network. Nerve-like neurons 6 and 2 of the east sample-A numerical identification system using neural-like networks, a set of target values in a shirt image. The identification system includes: knife ~ an operating device for Capture the image; First, change the device for converting the image into a grayscale format; Niche device is for positioning the partial image of the value according to the converted image; ^ A selection device for selecting the target value A specific image in some images 'The specific image contains a specific target value T of the set of target values T a normalization device for converting the specific image into a normalized image array'; a type of neural network, the normalized image array Input to this type of neural network 591545 6. Apply for patents for calculation and analysis, and get a set of analysis values; and a reading device, which reads out the specific target value based on the set of analysis values. 6 3. The numerical identification system as described in item 62 of the scope of patent application, wherein the numerical identification system further comprises: a determination selection device for judging whether the partial image containing the target value has been completely selected by the selection device. 6 4. The numerical identification system according to item 62 of the scope of patent application, wherein the numerical identification system further comprises: a storage device for storing the specific image after conversion. 6 5. The numerical identification system according to item 62 of the scope of patent application, wherein the positioning device further comprises: a marginalization device for finding an edge portion of the converted image; a level statistics device for horizontal accumulation A pixel of the edge part obtains a horizontal statistical result; a vertical statistics device for vertically accumulating pixels of the edge part to obtain a vertical statistical result; a horizontal positioning device 'finds the target value-containing portion according to the horizontal statistical result' A horizontal area range of the image; a vertical positioning device, based on the vertical statistical result, finding a vertical area range of the partial image containing the target value; and Page 591 545 A demarcation device analyzes the horizontal area range π 乂 and the vertical mu ρ * μ to determine the fan garden of the partial image containing the target value. The range of the domain 6 6 · As described in the patent application scope 6 2 ^^ τ strokes a few prostitution, 隹 feng a ba · t identification system, complex normalization device further includes A first device for normalizing a standard array of & numbers for a specific image; and a second device for normalizing the standard array to be at least one normalized image array with a fixed row. 67. The number described in item 66 of the scope of patent application: Her 4 / | value identification system The fixed number of columns is 21 columns. Wherein, according to the numerical identification system described in item 66 of the scope of the patent application, the number of rows of the standard array is determined based on the range of the part image containing the target numerical value ^ and determined according to a predetermined ratio. The dy-type neural network includes a round-in layer, a layer, and an output layer, the first hidden layer, a first regional receptive field exit layer, the input layer has a first regional receptive field, and the first: the hidden layer has a The second regional receptive field, the second hidden layer and the third regional receptive field, the type of neural network further includes: a, for sequentially inputting the normalized image array to the input layer f the first scanning device According to the first regional receptive field, scan the image array input to the input layer, and output to the first risk—a hidden layer 591545 6. Patent application scope A second scanning device that scans the image array input to the first hidden layer and outputs to the second hidden layer according to the second regional receptive field; and a third scanning device according to The third regional receptive field is scanned into the image array of the second hidden layer and output to the output layer to obtain the set of analysis values. 70. The numerical identification system as described in item 62 of the scope of the patent application, wherein the reading device further comprises: a selection device for comparing the set of analysis values to select a maximum value of the set of analysis values; and The judging means judges and reads the specific target value based on the maximum value. 71. The numerical identification system according to item 62 of the application, wherein the neural network further includes a plurality of neural-like sub-networks with different training samples. 7 2. The numerical identification system according to item 62 of the scope of patent application, wherein the image is a reading image on a meter device. 7 3. The numerical identification system according to item 72 of the scope of patent application, wherein the meter device is a gas meter. 7 4. The numerical identification system according to item 72 of the scope of patent application, wherein the meter device is a water meter. Page 53 591545 6. Scope of patent application 7 5. The numerical identification system described in item 72 of the scope of patent application, wherein the meter device is an electric meter. 7 6. A long-distance numerical identification system using a neural network, which is used to identify a set of target values in an image and send it to a data processing device at the remote end. The image contains a partial image containing the target value. The long-distance numerical identification system includes: an acquisition device for acquiring the image; a conversion device for converting the image into a grayscale format; a positioning device for positioning the target value according to the converted image A partial image; a selection device for selecting a specific image in the partial image including the target value, the specific image including a specific target value in the group of target values; a normalization device for converting the specific image into a normalized image Array; a type of neural network, the normalized image array is input to the type of neural network for calculation and analysis, and a set of analysis values is obtained; a reading device judges and reads the specific target value according to the set of analysis values; And a transmitting device for transmitting the specific target value to the data processing device. 7 7. The long-distance numerical identification system described in item 76 of the scope of patent application, wherein the long-distance numerical identification system further includes: P.54 591545 VI. Scope of Patent Application 1. Determine the selection device 'for judging whether the part of the image containing the target value has been completely selected by the selection device. 7 8 · The long-distance numerical identification season described in item 76 of the scope of patent application, the numerical identification system further includes: 充, a storage device for storing the specific image after conversion. 79 Medium • The long-distance numerical identification system described in item 76 of the scope of the patent application, the positioning device further includes: /, a marginalization device for finding one edge of the image after conversion · the value of the droop value is 80 = Flat statistical device for horizontally accumulating the edge part; to the flat statistical result; I serve a pixel for vertical accumulation of the edge part to get the straight statistical result of the vertical statistical device; horizontal positioning device, based on the horizontal statistical part of the image One horizontal area range; If the input contains the number of targets, the vertical positioning device will divide the vertical range of the image according to the vertical statistics; and the number of horizontal ranges of the target number and the range of the vertical range, as well as the numerical range of the three-dimensional range of images. The long-distance normalization device as described in item 76 of the scope of patent application further includes: a numerical identification system, which has a fixed row Page 55-Device for standardizing the specific image 591545 6. Standard array for patent application number; and a second device for normalizing the standard array to a normalized image array. 81. The long-distance numerical identification system described in item 76 of the scope of patent application, wherein the neural network includes an input layer, a first hidden layer, a second hidden layer, and an output layer, the input layer Have a first regional experience Field, the first hidden layer has a second regional receptive field, the second hidden layer has a third regional receptive field, and the neural network further includes: an input device for sequentially inputting the normalized image array To the input layer; a first scanning device that scans an image array input to the input layer and outputs to the first hidden layer according to the first regional receptive field; a second scanning device according to the second area The sexy receiving field scans the image array input to the first hidden layer and outputs it to the second hidden layer; and a third scanning device scans the input to the second hidden layer according to the third regional receptive field. The image array is output to the output layer to obtain the set of analysis values. 82. The long-distance numerical identification system described in item 76 of the scope of patent application, wherein the interpretation device further comprises: a selection device for comparing the set of analysis values to select a maximum value of the set of analysis values; and The judging means judges and reads the specific target value based on the maximum value. 8 3. The long-distance numerical identification system described in item 76 of the scope of patent application, which Page 56 591545 6. In the scope of patent application, this type of neural network further includes several neural-like sub-networks with different training samples. 84. The remote meter reading identification method described in item 76 of the scope of patent application, wherein the transmitting device refers to a wireless network system, the specific target number. The value is at least one of the wireless network system Data channel transmission. Page 57