Car Evaluation Using Neural Network Essay Sample free essay sample

1. Introduction Handwriting acknowledgment is done in two different ways. The first is online acknowledgment which examines the characters as the user is pulling them. This method is the simpler of the two. since the system lone trades with one character at a clip. An illustration of this method is character acknowledgment on a personal digital helper ( PDA ) . The 2nd type is off-line acknowledgment. In off-line acknowledgment the system must look at an full group of characters alternatively of merely one at a clip. An illustration of this is optical character acknowledgment ( OCR ) package for scanners. This system will utilize off-line character acknowledgment. Once the system has broken a image into its single characters. a nervous web will be used to find each single character. Following these characters. every bit good as information sing their locations. are sent to the scanner. The scanner so rebuilds the single characters into Numberss and besides determines which symbol goe s to the parser following. In some instances. the scanner must besides infix extra characters. The parser so requests one character at a clip from the scanner and calculates the look. Finally. a pop-up is displayed with the deliberate reply. Figure 1: Example 2. Images In this system. images are able to be input in two different ways. In either instance. images are required to be grey graduated table. Support may finally be added for non-gray graduated table images. but this was non considered of import for the initial version of the system. The first method of image input is with a electronic image file. The functionality for lading electronic image files was included for several grounds. First. since electronic image files do non compact the image data no external libraries were required. Therefore. change overing the file into a information construction used by this system was much simpler. Second. for proving the system. it is much easier to direct it a list of electronic image images to cipher instead than utilizing the graphical user interface ( GUI ) of the system to pull trial equations repeatedly. Finally. a future end of the system is to let users to lade images in from a scanner. so being able to manage image files will let thi s to work much more easy. The system presently does non back up lading images from a scanner because scanned images typically have a batch of noise ; in trials performed. this noise caused jobs when interrupting up the image into single characters. Figure 2: Noise in a scanned image This functionality will be implemented at a ulterior clip. However. the system will hold to filtrate these images and clean up the noise ( most probably by utilizing a Gaussian filter ) . and this was merely non executable given the limited clip restraints. It is besides a hereafter program to include support for other file formats of images ( JPEG. GIF. PNG. etc. ) . The 2nd method involved â€Å"drawing† the images on the screen through the program’s GUI. This method is used in the current execution. since it was considered the easiest and fastest for a user. Images are drawn by snaping and dragging the pointer around the draw panel of the GUI. Erasing is besides allowed utilizing the same method. The user may besides unclutter the full panel. When the user is finished composing an look he or she merely clicks on the â€Å"Calculate† button. The system so draws an lineation around each character it finds and displays a pop-up incorporating the deliberate reply Figure 3: Fictional character interrupt up For the undermentioned account. refer to Figure 3 for a graphical illustration of each measure of the procedure. Once an image has been loaded in the system ( Step 1 ) . it must be broken up into single characters. Presently. the system checks pels from left to right until it finds a pixel value below some threshold ( a black pel has a value of 0. and a white pel has a value of 255 ) . The system so creates a little bounding box around this pel ( Step 2 ) . Each of the four sides of this bounding box is checked to see if it crosses any pel below this threshold value. If it does. the box is extended in that way. This procedure is repeated for each side of the box until the borders of the bounding box cross no pels below the threshold ( Step 3 ) . This method works in merely some instances. since it is common that this bounded box will incorporate multiple characters. Some illustrations of this state of affairs include characters underneath a square root and bounds of an built-in. To take these excess characters. the delimited group of characters is scanned in the same manner from different waies. After a character is removed from the bounded group of characters. the group is scanned once more until no more characters are removed ( Step 4 ) . Finally. the jumping box of the original character is recreated since remotion of characters may hold affected its size ( Step 5 ) . This method has many defects. It is really successful in interrupting up characters that are non connected. but it is unable to interrupt up characters that are connected ( for illustration cursive authorship ) . Fortunately in mathematical looks connected characters are uncommon. particularly when composing on a computing machine screen. Hence. for the current undertaking this method was considered acceptable. Once the image is broken up into its single characters. each character’s location information is stored along with the pel values inside its bounding box. These pixel values are converted into a 10 pel by 10 pixel representation of the character. since the nervous web must be given a fixed figure of input pels for all characters. One job that arose with this method was that some characters. when converted to a 10 pel by 10 pels representation all look the same. For illustration. a really consecutive 1 or subtraction ( ) will turn into a block of all dark pels. and the system will be unable to separate these from a generation mark or denary point (  · ) . Besides. a somewhat slanted 1 will look a batch like a division mark ( / ) . To cover with this job. images that are really tall and narrow are padded on the sides with white pels. and images that are really broad and short are padded on the top and bottom with white pels. Figure 4: 10?10 jobs 3. Nervous Network The nervous web used for the acknowledgment of single characters is a feed-forward nervous web with four beds. The first bed contains 100 inputs. that is. one for each input pel. The end product bed contains an end product for each character that is to be recognizable by the system. Valuess for each input pel are sent into a corresponding node in the first ( input ) bed. For each node in the first bed. its input value is sent to an activation map. in this instance the logistic sigmoid function1. The end product of this map is sent to each node in the following bed. However. the end product it is non sent straight ; each end product is multiplied by some weight before traveling to the nodes in the following bed. Each node in the following bed amounts all of the signals it receives and sends this value to its activation map. This procedure repeats until the concluding end product vector to the web is found. x?= 1 1?e?t 7 For illustration. for the nervous web in Figure 5. to cipher the end product of node n+2. each end product for the old bed ( nodes 2 through n+1 ) must be calculated and multiplied by the corresponding connexion weight. This computation can be represented by the undermentioned equation: o n?2=? n?2 n?1 k=2 tungsten n?2. k o K ? O K is the end product of node K. ? K is the activation map of node K. and w K. J is the weight traveling from node J to node K. Figure 5: Sample unreal nervous web To develop the nervous web to acknowledge an individual’s script. a preparation set is created that contains a 10 images of each character the system is to acknowledge. The system can execute rather good when trained with fewer than 10 illustrations of each character ; nevertheless. 10 was chosen to guarantee a high degree of truth. The user has the ability to develop with more or less than 10 of each. but 10 is the default and the recommended sum of each. Each point in this preparation set is paired with a desired end product vector. This is basically a 0 vector except for the n-th component. which corresponds with this character. contains a 1. Next. the gradient descent larning method is used to develop the nervous web. Training is done by seting the weights of the web until the entire mistake for the preparation set is below. 005 a 2 B. where a is the figure of characters being recognized by the system ( figure of end products for the nervous web ) and B is the figure of each character in the preparation set. This can besides be thought of as the entire figure of end products when all inputs from the preparation set are sent into the nervous web multiplied by. 005. The entire mistake is calculated by directing each member of the preparation set through the web and ciphering the amount of the absolute values of the differences of the single constituents of the end product vector and desired end product vector. Weight accommodations are calculated with the undermentioned equation: w J. i=w J. one J oik ? is the larning rate ( in this plan this is merely 1 ) . and ? J is calculated as follows: K ‘ K ? J =e J ? ?net J ? if J is in the end product bed ? J =?’ ? net kj ? ? wm. J ? N if J is in a concealed bed m k net J is the amount of the inputs to node J for the K th component of the preparation set. Once the preparation is complete the weights are stored to a file which can sobe loaded in by the user instead than holding to retrain the system each clip it is used. 4. Scanner The scanner for this undertaking works rather otherwise than a scanner for a programming linguistic communication compiler. Normally. the following character in the sequence is the following character in the file ; nevertheless. in this execution the following character is non needfully known. When a user draws an equation. the system breaks the image into single characters and has the nervous web acknowledge these single characters. Once each single character is recognized and its location information is stored. this information is sent to the scanner. The scanner turns this information into items which are so sent to the parser. Numbers ( 0-9 ) and decimals ( . ) must be put together to organize the figure they constitute. For illustration. if the user writes the figure 10. 4. the system will see each character individually and must find that these four person characters make up the existent figure 10. 4. Besides. when two next footings are multiplied. for illustration 3x. the scanner must set a generation symbol between the footings. Similarly when the system encounters a power. it must infix the ^ symbol. so that the parser knows it has reached a power. When series or integrals are found. the system must look for bounds alternatively of merely catching the following character. The system must besides find which character or group of characters is the following to be sent to the parser. 5. Parser The parser takes the items given by the scanner and calculates the consequence. The parser must guarantee that operations are calculated in the right order and that operations that require numerical methods are calculated accurately. To cipher integrals. the trapezoidal method is used. The figure of trapezoids used additions until the estimated mistake is below. 00001 or until 100 trapezoids are used. Since computation of multiple integrals requires intense calculation. the fillet standards are relaxed in order to give a timely consequence. Figure 6: Example of trapezoidal method ciphering an built-in from 1 to 5 with four trapezoids 6. GUI The graphical user interface has a canvas for the user to pull on and allows the user to pull. erase. and clear the full canvas. The system has an option for making a new user. which allows the user to take which groups of characters he or she wants the system to be able to acknowledge. The system so asks the user to pull 10 illustrations of each character he or she asked the system to be able to acknowledge. Once all of these illustrations have been drawn. the preparation set is created and used to develop the new user’s nervous web. Once the new user’s nervous web has been trained. the nervous web weights are saved so that the user can lade his or her profile at any clip. 7. Future Work In add-on to the hereafter ends mentioned antecedently in the paper. one future end for the system is to include more mistake handling. Currently no message is displayed when an look is written or recognized falsely. Finally the system will explicate why it could non cipher the look and will propose alterations to the user. Similarly. the answer pop-up will finally include a TeX representation of the user’s input look. This representation will let the user to guarantee that the system recognized each character right. Another characteristic that will be implemented is the ability to add extra character sets to be recognizable by the user’s nervous web. That manner. if the user all of a sudden decides he or she would wish acknowledgment for integrals but he or she has non trained the system for this. instead than holding to travel through the full preparation procedure once more. the user can merely pull the new characters. The system will retrain itself utilizing this new informations and the preparation set the user antecedently created. This retraining will necessitate restructuring of the save file. since the original preparation set will necessitate to be saved. Similarly. a invitee user will finally be added that will be trained with all of the other users’ developing sets. This will let a user to utilize the system without preparation ; nevertheless. it will non acknowledge his or her authorship every bit good as it would if he or she were to travel through the preparation proc edure. Support is besides planned for ciphering derived functions. basic matrix operations. merchandise series. and roots of equations. 8. Decision For this undertaking I created a package system that allows users to handwrite mathematical looks which are so recognized and calculated by the system. This system is non yet complete. A working version has been created. but there are many characteristics that have non been implemented yet. Additionally. as with most pieces of package there are many bugs that need to be found and corrected. I plan to go on to work on this undertaking until all desired characteristics are implemented. Mentions 1. Stevenson. Charles F. . 1966. Neurophysiology: A Primer. John Wiley A ; Sons. Inc. 2. Gerald. Curtis F. and Wheatley. Patrick O. . 1999. Applied Numeric Analysis. 6th Ed. . Addison-Wesley 3. Russel. Stuart J. and Norvig. Peter. 2003. Articial Intelligence: A Modern Approach. 2nd Ed. . Prentice Hall 4. Schalko. Robert J. . 1997. Articial Neural Networks. McGraw-Hill 5. Li. Hongzing. Chen. Philip C. L. and Huang. Han-Pang. 2001. Fuzzy Neural Intelligent Systems. CRC Press LLC 6. Jang. J. -S. R. . Sun. C. -T. and Mizutani. E. . 1997. Neuro-Fuzzy and Soft Computing: A Computational Approach to Learning and Machine Intelligence. Prentice Hall 7. Kosko. Bart. 1992. Nervous Networks and Fuzzy Systems: A Dynamic Systems Approach to Machine Intelligence. Prentice Hall 8. Mammone. Richard J. and Zeevi. Yehoshua. 1991. Nervous Networks: Theory and Applications. Academic Press. Inc. 9. Principe. Jose C. . Euliano. Neil R. . and Lefebvre. W. Curt. 2000. Neural and Adaptive Systems: Fundamenta lss Through Simulations. John Wiley A ; Sons. Inc. 10. Foley. James D. . new wave Dam. Andries. Feiner. Steven K. . Hughes. John F. . 1996. Computer Graphics: Principles and Practice: Second Edition in C. Addison Wesley