TW578117B - Apparatus and method of enhancing image resolution - Google Patents

Abstract

A kind of apparatus and method for enhancing image resolution are disclosed in the present invention. In the invention, the technique based on fuzzy system, which simulates human vision characteristic, and neural network, is used in the interpolating technique for the digital image. After a new original image is inputted, it is divided and sorted by image analysis module such that different kinds of image are processed individually by linear interpolation and neural network technique. Because the fuzzy system is formed according to human vision characteristic and the neural network is obtained after learning the real nature image, the vision effect of the image, which is processed and amplified by the invention, is extremely close to the real nature image.

Description

V. Description of the invention (1) Field of the invention: The relevant device can be enhanced image resolution (reSGluti〇n) = set and method 'system is about-a kind of use simulation ^ η) (human vision excellent feature network ( ne 丨 丨 /) based on fuzzy analysis and neural-like technology to enhance image resolution. BACKGROUND OF THE INVENTION: With the rapid development of technology, digital imaging products are very popular. For example, digital cameras, digital Cameras, projectors, and multi-function business history. However, all digital imaging products have a common technical requirement, which is the adjustment or conversion of image resolution. Therefore, high-quality image resolution enhancement technology directly addresses an important issue in digital image processing. The so-called image interpolation (interpolation) technology refers to converting a low-resolution image to a high-resolution image by interpolation. However, common image interpolation techniques, such as Bilinear interpolation (Bilinear interpolation) and Bicubic interpolation methods are mostly unable to meet the needs. This is because of the conventional image interpolation techniques. There are two obvious shortcomings such as jaggedness and blurring of the image. 'There are many interpolation techniques based on the linear interpolation principle, and most of them also have this disadvantage. Because of the human visual system's Sensitivity The edge part of the image will be sharper than the smooth or structured part of the image. Therefore, the current 578117 V. Description of the invention (2) There are many technologies that take special consideration into the edge of the image when processing image interpolation. In part, in recent years, the concept of image contour has been gradually introduced into the problem of image resolution enhancement. For example, Pohsiang Hsu et al., US Patent 5,991,464, "Method and System for Adaptive Video Image Resolution Enhancement" ", Is very representative. Later, the interpolation technology of digital images based on contours and edges in the image has become a development trend, and highlighting the edges in the image to facilitate human eyes observation has also become an important issue. ;by

Chien-Hsiu Huang in the 2001 U.S. Patent [^ 6,175,659 "Method and Apparatus for Image Scaling Using,

“Adaptive Edge Enhancement” is an image interpolation technology based on enhancing the edge of an image. In the US Patent Case PUB · APP · NO · 20 020 1 2690 0 proposed by Kim and Sang Ye on in 2000, the "Image Interpolation Method and Apparatus Thereof" combines traditional interpolation technology and the direction of the edge of the image. As a way to deal with it.

However, how to design better interpolation techniques and how to evaluate the pros and cons of processing have always been challenges for the development of related technologies. Therefore, the present invention proposes to use the ability of neural network-like self-learning as a design basis for interpolation compensation. In addition, the present invention also uses the concept of human visual characteristics to design a set of fuzzy systems for image analysis. System as a device for image classification 'simultaneously combines bilinear interpolation and neural network-like interpolation methods

r〇, "I use’ to achieve a balance in image quality and processing time, and can obtain better quality than the image produced by the conventional magnification technology, whether it is on the side of the image

^ / «117 V. Description of the invention (3) The sharpness of the edge or the hard work of the image are more excellent. The purpose and overview of the invention ... The main purpose of the present invention is to enhance the resolution of the image via the network for a use The digital analysis of fuzzy analysis and god-like input and $ 2 placement, this method can make the effect of the natural image of the user, and make the image magnification of the shirt very close. The visual effects and actual natural images are very close to the other of the present invention. The device and method are based on simulation, analysis of 2: 2 :: enhanced image resolution, and neural network-like characteristics. The fuzzy analysis system enters the resolution of the digital image; one way is to implement the natural nature of the image: the type of the image, and to carry out the invention-the purpose of the fuzzy analysis system is to use a simulation of human visual characteristics; interpolation or neural network Interpolation on the road: Secondly, in order to choose to use the bilinear "in the image quality and process, the mechanism of processing digital images, so as to achieve a balance between the present invention and the pi bang. A special neural network-like framework is designed to design a set of training features, and the actual image of this network framework is like the actual natural :: = method, which makes the invention of the present invention "rare. 3 and methods, its image resolution The degree of conversion and the smoothness of the image are both better. Page 6 578117 V. Description of the invention (4) ^ To achieve the above purpose, the present invention first obtains an original image A and then the fruit paste analysis system Analysis of raw images It will have edge characteristics and no edge characteristics; then: 箪 1: class is a region with edge characteristics' use-angle meter 捭.: After the household edge angle, use neural network-like interpolation technology to directly benefit the image The internal technology, which is classified as a region-free image without edge characteristics, performs image processing; after finishing finishing, you can get-high-resolution digital images. It is easy to explain the specific embodiments with the accompanying drawings: easy to understand this The purpose, technical content, characteristics and features of the invention: Description of drawing number: 10 Low-resolution original image, and then the original is in the original computing module image area, then the original 0, when the success is 14 18 22 26 30 34 40 44 46 48 Classification module High-resolution digital image angle calculation module Bilinear interpolation module Image mask position to be compensated Pixel-like neural network Hidden layer Hidden layer Output layer Neuron image analysis model Group of image interpolation modules fuzzy analysis system neural network interpolation model database original image (grey part black part) 42 input layer 442 neuron 462 neuron 12 16 20 24 28 32 Group

578117

Detailed explanation: ^, Ming proposed a method to enhance the natural image resolution and image rule # 么 #, — Wanfa, which uses the fuzzy points that simulate human visual characteristics, (ei, perform. Shirt image analysis' and use objects in the image The direction of the edges is based on neural network-like technology and is used for digital image interpolation. = As shown in the first picture, after a low-resolution original image 10 is input, = an image analysis module 12 performs image segmentation and analysis, and a classification ^ group 14 is based on the analysis result of the image analysis module 12, The original image 10 knife = smooth / structure area or edge area; finally, through an image, j-mode and 16 are used for scene image processing, which uses the bilinear interpolation technology to the original shirt image Zhi = slip / structure area for image processing, and use neural network-like interpolation technology for image processing on the edge area of the original image to obtain a resolution digital image 18 output. The second figure is a schematic diagram of the architecture of the present invention using a fuzzy analysis system. As shown in the figure, S — the original image j 〇 is input into a fuzzy analysis system 2 0, and the fuzzy analysis system 20 starts analyzing the original image 丨 0. , And classify the original image 10 as having edge characteristics and not having edge characteristics, so as to distinguish the image 10 as whether it should use the bilinear interpolation technique or the neural network interpolation technique. The original image 丨 0 is classified as a region with edge characteristics, and then an angle calculation module 22 is used to calculate the edge angle, and then the neural network-like interpolation module 24 is used to perform image processing on the original image 10; If the original image 10 is classified as a region having no edge characteristics, the original image 10 is directly subjected to image processing by using the bilinear interpolation module 26. original

Page 8 578117 V. Description of the invention (6) = After the processing described in ii, the digital η with enhanced resolution can be obtained. This type of neural network interpolation module 24 uses a supervised method to train a neural network. The relevant parameters of this type of neural network training are stored in a database 28. After referring to the image processing process of the present invention, please refer to step S10. An image is used in the step S10. Segmentation of the image to facilitate subsequent analysis steps 丨 Another fuzzy analysis system analyzes and judges the original image. As shown in the steps, the original image is classified as having edge characteristics and no edge characteristics. Output variable (M〇) Prediction: Threshold value (Th), then the pixels in this area have edges = select this type of neural network interpolation technology, and perform step s "differentiate its edges, and angles. If step S16 uses neural network-like interpolation technology, the original image is processed by the image. If the output variable is smaller than the critical value, the pixel in this area does not have edge characteristics. This means that step should be selected. double The linear interpolation technology performs image processing on the original image, as in step 32, to obtain the processed image pixels, and then, as in the step, determine whether the entire image of the original image has been completely processed. If not, then If the image mask is moved to the next processing position in step S24, repeat steps S12 to S2. If the entire image has been completely processed, μ P can serve the high-resolution digits processed as shown in step S26. Image, so that the entire image processing flow can be ended. F ^ 2 The present invention can also use a conventional image edge extraction module to first extract the edge portion of the "α shirt image, and then use an angle meter 578117. 7. Description of the invention (7) The calculation module calculates each angle and edge position information on the edge of the image and analyzes it. == Divides the image into a bilinear interpolation technique: the original shadow. Yi 11 X 疋 uses a neural network-like interpolation technique when the original image enters the image blurry image. The image is masked; it will be covered by this image mask. The second cut is shown in the fourth picture, and in the image, the gray part 32 in the picture is the inner 34 of the original image, which is the position of the compensation pixels required for the pre-image enlargement. # & $ a, μ pixel value represents the original image, _ stands for reference pixel (〇M ,, ^ ^ ^ t nelghb〇rh〇〇d pixel), " Pixel values. Since the image mask 30 will be shifted in the original =, each image value to be compensated will be determined, and ^ = the pixels interpolated on the image ... from the formula ⑴; A pixel has a weight value of 0 (i J), and% ... is obtained by stealing the neural network designed by the present invention.疋 利 ρ —) = ΣΣ, H ·) i-0 j-0. (1) Due to the physical limitations of the conventional image interpolation technology, blurs and jagged edges in 3 contours and edges are generated. Image, because the quality of the image is low, an ideal magnification technique should consider the directionality of the edges in the image, so that a clearer and sharper image contour can be obtained.

Page 10:) / 0 丄 丄 / 5. Description of the invention (8) In order to achieve this goal, the characteristics of the viSUal system of the present invention are used as visual systems (Hu called the human visual system. Mechanism, the reason is that the characteristics of the human eye can be used to analyze; j points have a more sensitive perception, and the eight-body principle friction examples to verify the description ^ This technology will be able to refer to this ^ intraoperative valley and make the heat The practice is to implement. ^ The stroke of the example has acquired sufficient knowledge and according to the fuzzy analysis system: Φ two perception characteristics, we can see that the human eye for the image, [, t is more than the perception of the degree. Keen

The difference between the object and the background in the eye image is the average value of the two squares, as shown in the fifth figure. The vertical axis in the figure represents the critical value of human visibility (visibiHty thresh〇id). The horizontal axis represents the average background brightness (BL) in the image. It can be seen from the figure that when the average background brightness is in very dark and bright areas, the threshold of visibility ^ becomes high, which means that human vision is The force of discriminating objects and background in these two areas becomes worse, and the color (or brightness) of objects in this area and the background must be very different in order to be easily discerned by the human eye. Conversely, when the average background shell degree falls between 70 and 150, the difference between the human eye and the background can be easily distinguished. In other words, the objects in this area and their backgrounds are in color. (Or brightness) doesn't need to be very different. In addition to the ability to identify objects and backgrounds, the human visual system will also respond differently to images of different structures. The human eye responds to high IHI surfaces. Page 11 578117 5: Description of the invention (9) ^ —- The range of contrast Such as structure, image contours, and edges all react, but not for smoother image areas. =; The same; = the quality of the image after processing and the fast processing time, the present invention proposes a human vision system-based fuzzy analysis system, which ^ = = T divides the input image content into having edge characteristics and not having 纟There are two types of force crimes, and the part with edge characteristics in the image will be beneficial; the nerve :: sexual interpolation technology will be used for processing, and the part without edge characteristics will be processed with large techniques such as recording, and coordinated with parameter adjustment. In order to make the best arrangement for the quality of the image quality and processing speed. It can be seen that in the framework of the fuzzy analysis system of the present invention, there are = for visibility calls, structural calls, and ㈡ = numbers. The variable (Mo) is the selected mode. As in formula (2), and the wheel V (BL) = 20.66e ^ 3BL + e0008BL, it is shown that (2) uses non-linear regression to find the .V (BL) in the fifth figure, where BL Represents the background in the N * N image mask 妁 =: The difference between D and V (BL) can be used = thousands of degrees of freedom VD = DV (BL), formula (3) '

Among them, D is the maximum pixel H of the original image (4), the difference between m Jing, such as common D = max (〇 (ij)) 0 (ij)) Formula (3) indicates that the light is produced in a certain background (4 ) Is there any object visible to the human eye, if the VD " 'N * N image masked objects and background are easier to distinguish, otherwise # |, two means that the right of the range is less than zero, which means that the Fanyuan

578117

For non-visually sensitive areas. In addition, SD and CD are used to analyze whether the pixels in the mask have the characteristics of contour edges. SD reflects whether there are obvious two groups of characteristics in the image mask, as shown in formula (5), max (, y))-Σ J) + J)) SD = _ _ m ^ (〇 (U ·))-min (0 (U ·))) (5)

If SD obtains a large value, it means that the image mask does not have the characteristics of two groups. At the same time, it may be noise or a smooth area. If the fairy value approaches zero, it means the image mask. There are two groups of characteristics, including those that can include the edge or structure of the image. Since these two characteristics will show the same value after other operations, the variable CD is added here, as shown in equation (6) Shown

NN 1 (6) Here, the present invention uses the concept that differential is very sensitive to noise to analyze the types in the image mask. According to formula (6), it can be seen that if the image mask contains the structure type, The CD value will be higher, but if its content contains images with edges, the CD value will be lower.

The visibility VD of the input variable contains two fuzzy sets (negative (N) and positive (P), respectively), and the structural SD of the input variable contains three fuzzy sets, small (S) and medium (M). And large (B), and the complexity of the input variable CD also includes three fuzzy sets S, M, and B '. The output variable Mo includes bilinear (bi 1 i near, BL) and god-like

Page 13 578117 V. Description of the invention (11) Two fuzzy sets via the network (neUral netwrk, NN), and the member functions referenced by each variable are as shown in Figures 6 (a) to 6 (6). The fuzzy set described by i and the member functions used will make various changes according to the different uses of f; and the following seven fuzzy rules are based on the variables and member functions mentioned above. Examples: · If VD = N, then M0 = bl; If SD = B, then; Right CD — B, then M0 = BL ·; If VD = P If VD = P If VD = P If VD = P Where SD = s and CD = S, then Mo = NN;, SD = S and CD = M, then Mo = NN;, SD = M and CD = S, then Mo = NN; and SD = M and CD = M, then Mo = BL. The output variable Mo will be prepared after defuzzification: ι = 0 ~ 1 〇, the * character, ☆ This invention can deal with a critical value (Th) according to the bite, when the rib gTh, the present invention, As if via Internet interpolation technology, on the other hand, if Mo < Th, a double ^ type god technology is selected for image processing. Laughing linear interpolation In terms of angle calculation and neural network-like interpolation technology: In recent years, many shadows based on the contour direction of the image have been proposed. 'However, how to design the most appropriate shadow interpolation technology and how to judge the processing The quality of the image is an eighteenth street. The present invention is based on a combination of bilinear interpolation; a difficult interpolation technique to solve this problem. When the internal network analysis system of the image mask is classified as having the characteristics of edges, the angular sound of the edges = fuzzing & will be counted 578117 V. Description of the invention (12) For calculation, please refer to the third figure As shown, 邶, the force is expressed as the angle of 0 (~) in the image mask at this time. The calculation method of other forces is shown in formula (7), =-η fr.--[Tan (7) (8) this The force a at the counter a, j) is as shown in formulas (8) and (9)-(0 (i +1, j — 1) + 20 (i + 1, force + 〇 (Σ · + 1 "" + 1)) Do = — U — l) +20 (i, j-l) + 〇 (i + U— 1)-(0 (i -1,) + 1) + 20 (ί, ”+ 1) + 〇 (ί + ι, j + 1)) (9) In the above formula, the range of i and j is AND, and the resulting angle can be quantified to the required division according to the requirements. If the reference pixel is classified as non-directional, a bilinear interpolation technique will be applied, as shown in formula (10), to process the pixels that need to be interpolated to reduce the amount of computation, 〇 (U)

ΣΣ—_two one V (m7n) = ^ M Stance (0 (iJ \ P (m9n)) ΣΣ ~ — ^ _ μ / -1 distance (〇 (ifj) 9P (mrK))

c ^ (10) / Conversely, if the reference pixel is classified as having a major directional characteristic, the system will use neural network-like interpolation technology to process it, considering the quality of the human visual system. The present invention uses a supervised learning rule to train a neural network. First of all, the input and output expectations of the network must be obtained (desiRd), so that the input data can be used to enter the network to obtain consistent output values. Output expected value compared, the difference between the two

Page 15 578117 V. Description of the invention (13)-It will be the basis of the training network of the present invention. In detail, the present invention uses a high-level image capture device to obtain several high-resolution digital images. Assuming that a certain high-resolution image is ΙΝχΝ, the INxlT content is sampled and a new image is obtained. · MχΜ, where M = N / k and to do, 7) = edition, secret, that is to say 1 is the ideal image of ι · ΜχΜ after k times magnification, so you can get nerves in ι · ΜκΜ The input value of the network and the expected output value are obtained in it. The weight training process is shown in Figure 7. When a high-resolution original image 10 enters the neural network 4 0, the obtained limbs and high resolution Compare the content of the digital image 18, between the two :: will be used as the basis for the 4G weight of the modified neural network, and: # 8, the United States stores the correlation of this type of neural network 40 after training Reference = In the Ming system, the image does not have the edge contour feature in the linear interpolation technique. Therefore, it is used to train the neural network 2 to use the double material to "..." in the pixel area of the edge corridor: Zhong Xun 7: The training plan can learn that 'assuming the heart is The fourth reference in the image is _ = plus, and after the pixel is analyzed, it can be expressed as the… The path is I (ki + m, kj + «) 〇 孓 甲 3 and waiter, and said that the present invention is designed to be used as shown in the eighth figure, this network is a network architecture including — 声 = #, such as

On page 16, the input layer 42 and the second hidden layer 578117. The description of the invention (14) 44 46 and the output layer neuron 48 are four-layer neural networks. The input layer 4 2 and 3 of the network have two variables. They are $, claw, and n, and the second layer of the network is the hidden layer _44, which is the first layer of the hidden layer. It contains τM nerves = 442. According to experimental experience, the number of M is about 15. Between ~ 300, inter-applications are not limited to this; and in neuron 442, the present invention uses a bipolar function (bip〇lar ActivatiOn FunctiOn), as shown in formula (1 1) No, as an example of the function of the present invention (ad 丨 vati〇n fjnCt ^ 〇n), in addition, other function can be used according to different applications. A 1 l + e ', —the weight value between the round-in layer 42 and the hidden layer 44 in Zhou Road, which means that the second layer of f Erhuang Road is the hidden layer 46, which is the second hidden layer' Contains N * N neurons 4 β?, Η η junction ^ ^ ^ A: Wu 462 and also use the binary value function to do 4 6 Η; estimate the second and hidden layers of the network "and The kernel value vector between the hidden layer 46 of the third layer is represented as p, and the output of the hidden layer 462 is represented as ground) (^ 70 ^ this ^) represents the fourth layer of each network in formula (1). The output layer is a neuron's heart, which represents: interpolated pixel values ⑽), and the weight vector between the hidden layer and the neural network 48 is represented by the angle of the zeta angle ^ The table does not represent the pixel value of a certain ^ four solid; ^ ^ near 'pixels, please enter at the same time, as shown in formula (12). See arsenic-like additional input via the network 578117 V. Description of the invention (15) And the output of the network can be expressed as iW7 == Σχ ^) · 4 (12) (13) At the same time the relative expected output value is expressed as ^ = I (ki + (14), and the parameters, #and, and eZ7 are updated The method is as formula (15) And public (16), after the neural network-like network training is completed, this hair is' the corresponding neurons obtained through different inputs 462 ^ (¾) as a tool for A y intended natural image magnification ^ (^ + 1) = ^ y (Sa)), 2] xg (Zj ～ + 1) = ～ 纠 π Xin⑽―Xgi)) v 』jx [〇 + g (^)) (lg (Z ,)) / 2] / ^ The present invention selects several natural images as training samples, and sends (16) points including self-deficit shots using high-resolution digital imaging equipment ^: most = neural network-like , And then used the two natural shadows to complete the two objects, the flower bushes and the propeller. In this experiment, Shizheng D was used as the proof and test. The present invention interpolates the image (15) Page 18 578117 V. Description of the invention (16) The original image compared with the two traditional square interpolation methods, and the cubic interpolation method and the interpolation degree and image boundary proposed by Ming after the image is enlarged by 400%. Therefore, the present invention uses Afterwards, the “boundary fruit in the image is determined as a nerve” is described by the above-mentioned quality and techniques in the experimental paste analysis system, and the neural network-like shadow practice As a known method, use neural-like visual interface to connect the image or use the appropriate variable or balance in the bicubic interpolation of the object in the image. Within the linearity of the ninth (b) to (d) of the present invention The results of the sharpness produced by the technology are the same as the image components of the edge-wheel network image of the human visual image. The parameters of the network image interpolation technique of the present invention can be used for various parameters. The parameters of the results of the corpus perimeter are based on the interpolation technique map and the figure, respectively, and the experimental shadow ratios are based on the experiment. When looking for the magnification technique, you must know the knowledge and the time, combined with the fuzzy habit. Judging from the interpolation skill of the technology, the sharpness is excellent. It is in the shadows: the double-line ten pictures are visualized by using the network results. In order to achieve the analysis system, design the natural shadow foundation, there are many 'no matter where the technique is located, and like qualitative interpolation and dual (a) is the non-amplified front bilinear interpolation method, double interpolation technology, See the original video , Using the present invention is excellent in tube method of image clarity. The analysis system determines the characteristics of the edge contours in the person. In fact, the effect of the enlarged image on the final image can be adjusted by changing the model to adjust the image processing requirements. The completion of the unified measurement of the edge of the image is different from the previous image using the characteristics of science and the results obtained from the experimental results, based on the image's sharpness bilinear interpolation method or for different applications. Acquired in processing time

578117 V. Description of the invention (17) = The above-mentioned embodiment is only for the purpose of explaining this point. Its purpose is to familiarize yourself with this technique ^ technical heart and special content and implement it accordingly. It is enough to understand the limitations of the present invention ^ ^ ^ and the scope of the patent of the present invention, that is, equal changes and modifications made by the spirit disclosed by the big t ^ ^ should still be included in the scope of the patent of the present invention. iGo /% p. 20 578117

Brief description of the drawings Brief description of the drawings: The first diagram is a schematic block diagram of the device of the present invention. The second figure is a schematic diagram of an architecture using a fuzzy analysis system according to the present invention. The second figure is a schematic flowchart of image processing in the present invention. The fourth figure is a schematic diagram of an image mask used in the present invention. The fifth figure is a curve diagram of human visual characteristics. Figures 6 (a) to 6 (d) are the member function illustrations of the input variable visibility, structure and complexity, and output variables of the fuzzy analysis system of the present invention, and their intentions are shown. The seventh diagram is a neural network of the present invention and the like. The eighth diagram is the image used in the present invention. The ninth (a) to ninth (d) diagrams respectively illustrate the conventional bilinear interpolation method and the double result. Figures ten (a) to ten (d) are the results of the invention and the conventional bilinear interpolation method, respectively. Road weight training process diagram. Schematic diagram of neural network architecture such as interpolation. The original image of the flower bush and the magnification of 400% by the cubic interpolation method of the present invention The original image of the propeller and the 400% magnification by the cubic interpolation method of the original

Claims (1)

578117 VI. Application for Patent Scope 1. A device capable of enhancing image resolution, including an image analysis module, which captures and performs segmentation and analysis according to i;-the original image of resolution, and a classification model Group, which is classified into smooth / structured area or edge area according to the image, and analyzes the ^ Gift, and ^ I ΐ ^, 1 ± "# ^ ^ 丨 in the heart / month / L structure area Image processing is performed, and the interpolation technique is applied to the edge area of the original image to pass through the network-high-resolution digital image. Into the "image processing, and then get, 2 units, as described in Gan Shen ^ Patent Fan® item 1 can enhance the ^ image resolution device, where the image analysis module system includes: fetch the image and extract the edge Unit, which is used to set the edge region of the original image to be viscous3 to 4 and set it to a corner ^ calculation unit, which uses the angle represented by each pixel on the edge region that has a directional characteristic of the edge region. — Such as The device capable of enhancing image resolution as described in item 1 of the scope of the patent application, wherein the image analysis module and the classification module perform image analysis and classification by using an analysis system. The image interpolation module uses this type of neural network interpolation technology to image the marginal area of Xili and Detao before using the one-angle ^ count ^ two groups to calculate the angle of the marginal area. The device capable of enhancing image resolution as described in item 3 of the scope of patent application, wherein after the fuzzy system performs image analysis and classification, and in the 5, a method for enhancing image resolution, including the following steps Page 22 578117 6. The scope of the patent application provides an original image; analyze the original image, and classify the original ~ image knife as having edge characteristics and not having edge characteristics; i ϊ ΐ ΐ ΐ image is classified as having Areas with edge characteristics ', then Ye Lili's image ii' uses the neural network-like interpolation technique to the original image = the region in the original image classified as having no edge characteristics $ The original image is bilinearly interpolated Perform image processing: and process the entire original image 'to obtain a high-resolution digital image, as described in item 5, which can enhance the image resolution'-fM 象 ： ί = Before performing the analysis step, you can Use the second image masking army to segment the original image to facilitate the subsequent analysis. Step 8: Method to enhance the resolution of the image as described in item 6 of the scope of patent application — 'This image mask system Consists of pixels. In addition, the method for enhancing image analysis described in Item 5 of the scope of patent application for the application is that the fuzzy analysis system has visibility, structure, and complexity. There is also an input variable to classify according to the input variable. '9. As described in ^ Please refer to item 5 of the patent scope for enhancing image analysis. ΊYu = Fuzzy analysis system is performing classification, if it loses: variable system ΐ;:;: when preset threshold value' This type of neural network interpolation technique will be selected. When the output variable is smaller than the critical value, the interpolation technique performs image processing and selection on the original image. Hai Shuangxian [Health, page 23, patent application scope _____ f 所述 As described in item 9 of the patent scope ‘where today yesterdayxing # 〃 The resolution of the enhanced shirt image is too small. The digital image and conditions of the image are as described in Item 5 of the scope of the patent application, where the neural network interpolation technology is used in a powerful way; the square of the resolution = via the network, and the network The training method of Road 2 and I sub-provincial method is stored in a database. 1. Relevant parameters after mining. 2 A method that can enhance the resolution of the image, including the original image provided by π, including the following steps: Extract the edge part of the original image; calculate the image Each pixel on the edge region represents, according to the position and calculation of the edge region of the image, '= the original image is classified as using a class or 疋 bilinear interpolation technique; the interpolation technique uses this type of nerve The network interpolation technology and the bi-linear initiation image are used for image processing; and Yishu Shu completes the processing of the entire original image to obtain a high-resolution Lu Zhili image. X 1 3. The method for enhancing image analysis as described in item 12 of the scope of patent application, wherein before the step of capturing the edge region of the image, X ^ 1 uses an image mask to segment the original image. To facilitate subsequent analysis steps. 14. The method for enhancing image analysis as described in item 12 of the scope of patent application, wherein the image mask is composed of N * N pixels. Λ 1 5. The method that can enhance the image analysis as described in item 12 of the scope of patent application Page 24 578117 VI. Patent application scope method, in which this type of neural network interpolation technology uses a supervised learning method to train a neural network, and stores related parameters after training of this neural network in a data In the library. Page 25