人脸识别 眼睛嘴 轮廓都解决

所需积分/C币:31 2018-07-14 14:28:44 1023KB PDF
23
收藏 收藏
举报

人工智能 人脸识别所使用的方法大全,针对脸部轮廓 眼睛 鼻子 嘴巴
Face contour extraction 1l69 START PARAMETER COMBINATIONS DE PARAMETER ACCUMULATE WORE PARAMETER CELLS EDCE FOR AN EDGE PIXELS? TARGET FIND PEAK IN SEACH AREA IN PRECISION ACCUMULATOR RECHEDT DENTFY CURVE IN IMACE END Fig. 1. The simplified AHT algorithm operation of the vote Z of a parameter cell, i.e and shape of the neighborhood to change at each Z: =Z+E+C o iteration(i is not held constant), but we modify the where C is a constant. A small C value will emphasize algorithm is marked with an asterisk in Fig. 1. Since the weighting of the edge strength. In our experiments the accuracy of detection is mainly affected by the with both cheek and chin line detection, it was found quality of the ed ge image, our simplification is justified that setting C to 0 produced the most acceptable Also, using the original AhT may require more time results because the effects of noise edges could be re- The size and shape of the search neighborhood should duced be carefully selected to balance accuracy and computa tion speed. A smaller neighborhood allows the para 5. Neighborhood of a peak. Because of the discrete meter space to be more finely divided and can save nature of digital images and the noise involved in computation time, but a larger range will cover more peak in the accumulator array is not cases, increasing reliability. The choice of accumulator always exactly at the correct cell. A neighborhood of resolution and the size of peak neighborhood affect the the detected peak cell must be chosen in order to number of iterations k required to reach the target capture the" true"peak. USually, the neighborhood precision g. Consider a one-dimensional case. For the takes the shape of a multidimensional rectangle arameters xo and yo in the parabola, g is i pixel. Let x b m denote the number of cells the vided into, L denote the size of the search area, and b The AHT method proposed in Ref. 20 allows the size denote the neighborhood size. The number of required X. LI and N. roeder iterations k will be The value corresponding to the middle of the cell is considered to be the value for that cell (2) We now report our findings in the decision making log(b/m on the issues mentioned in section 2 6. Number of peaks. When more than one para- metric curve is of interest, more than one peak in the 1. Relevant subimage. The Hough transform and parameter space must be detected and refined further. the deformable template method(1) can provide the The number of curves to be detected differs from appli- positions of the centers of the two eye irises and the cation to application vertical position of the mouth, as shown in Fig. 3. Two rectangular regions of thc input imagc arc used. The End points of the result curve. When the target vertical search area covers from the pupil to the mouth precision is reached, a single curve in the image is and the horizontal range of the rectangle is between detected, which would be the end of processing in some l1w and l2w units away from the left left iris. That is, the applications. In our face contour extraction, we also width of the rectangle is relative to the distance w. The need to know the ending points of the curve in the search area for the right cheek is defined similarly. A image. The determination of the ending points is a conservative estimate was used initially to setl, througl non-trivial task and will receive further discussion. l4(based on 28 test cases )and the true values of the Is were recorded in each test case in order to facilitate the One purpose of our experiments was to identify the parameter setting in future studies. Table 1 reports above issues and decide on good compromises due to some statistics of the l values from the 81 test images the trade-offs involved The next two sections report some of the decisions made on these issues in the cases For example, when w=50 pixels, if 1=60% then of cheek and chin detection l, w=30 pixels 2. Parameter ranges. Since the detection of close to 3. CHEEK DETECTION vertical lines is necessary the range of a to be searched is from 80 to 100. The range r is set according to the This section outines the procedure for detecting relevant subimage. Since the range for A is so close to cheek lines using the eye and mouth locations pro- vertical we can assume that the minimum r value vided by the algorithim in Ref. I. A simplified AHT corresponds to a line going through the lower left hand algorithm was employed to detect approximately ver corner of the relevant subimage with A as 100. also tical lines. The results characterize the width, shape the maximum r value corresponds to a line going and orientation of the face and thus contribute to the through the lower right corner with an angle of 80,1.e identification of the individual max-I max-y'max cot(80 ))cos 80%+.max 3. 1. The equation and parameters sin(80) In a front-view image of a face, the cheek lines are rmin=(ymin -Xmax tan 10)cos 10 approximately straighT, and almost always close to These measurements are depicted in Fig4 coordinate equation A+y sin depicted in Fig. 2. The parameter space is two dimen sional, and the value of r is calculated for each a value <④ to determine the accumulator cell for each edge pixel Fig 3. Relevant subimages for cheek detection Table 1. Rccordcd parameter values for the cheek regions x cosa +y sin A l Extrem 66.7 40.0% 389%691% Mean 576%0 551% Fig. 2. Parameter definitions in straight etection Face contour extraction 1171 peak is eliminated from the process of finding the second peak. The accumulator cell with the second y highest vote count V2 is also examined. IfV2/i>cheeks for a constant he, this cell will also be 1009 further Using a cheek value higher than 0.4 produces 80° similar cheek detection results, but any lower may cause detection of extraneous, incorrect cheek lines. In lI 81 test cases used in our experiments, only one straight line was eventually detected for each side of the face, so searching for one peak should be enough in future stud i 7. End points of the result curve. Among the edge pixels associated with the detected curve the ones with the extreme x (or y) values would normally be con y s sidered as the end points, in the case of a straight line To find the endpoints we use this simple technique for each x; in subimage calculate Di range based on A range; Fig 4. Finding the range of r set endpoint, second edge within range set endpoint= second last edge within range 3. Resolution of the accumulator. The size of the accumulator was chosen to be 8 x8. In this applica- 3.2. Experimental results and discussion lator will cause each iteration to require more computation USing a smaller accumula The cheek detection procedure has been tested using ne number of iterations to grow signific- 81 face images, where 162 cheek lines are to be de antly because of the large range of each cell. Our tected. Some faces are relatively large compared to the experiments showed that the required precision in whole image, such as that in Fig. 5(e); while some are both the r and A parameters could often be reached in quite small, such as Fig. 5(b). The individuals are of the same number of iterations when using a square different races and skin tones. Some have a beard or accumulator. A rectangular accumulator would use glasses more iterations of the aht than necessary Accurately detected cheek lines are recorded as good", while those that were detected at an approximate 4. The use of edge strength and angle informa location are recorded as"marginal". The cheeks comple tion. The angle of each edge pixel is used to decide on tely missed or detected at the wrong location are re the A value of the corresponding line. The form of corded as"wrong equation 3)allows direct use of the edge angles. recall Figure 5 gives examples of the cases where the cheek the definition of A in Fig. 2, the edge pixels with angle detection was successful on both sides of the face.The the straight lines of the form detected cheek lines are marked by"opposite"straight lines. That is, the gray level g(x, y)of each pixel in these r= x cos at y sin a. lines is changed to a new value g (x, y), where In practice one more cell on either side of the MG if g(x, y)<(ma+MG)2 calculated angle cell is also accumulated when calcula g(x, y) ting the r values. This way, only a portion of the mig(x,y)≥(Ⅶm+Mc/2 accumulator is covered, reducing computation while In this experiment, the minimum gray level me is O maintaining reasonable accuracy. and the maximum gray level Mg is 255. This way the line will appear as a black line in a light area [e.g 5. Neighborhood of a peak. In our experiments, Fig. 5(e)], and appear as a white line in a dark area a 3 x 3 neighborhood surrounding the peak cell is [e.g. Fig. 5(a)], assuming significant neighborhood used. Once a peak is found, this 3 x 3 region will be gray level correlation. The presence of a beard in the zoomed in on for further accumulations cases of Fig. 5(a) and(d)did not affect the detection, nor did the glasses in Fig. 5(d) 6. Number of peaks. Initially, the procedure was set Examples of "marginal"and"wrong"fits are shown to detect at most two peaks in the parameter array for in Fig. 6. A(marginal, marginal) case is given in Fig. 6 each side of the face. In the accumulation process of (a)and is marked as(m, m). Similarly, (marginal, good) the first iteration, the region of the accumulator array and(marginal, wrong) examples are given in parts(b) with the highest yote count V, is marked and will be and c). Parts(d)and(e), give two cases where one searched further. The 3 x 3 neighborhood of the first cheek line is detected correctly and the other one is at X li andN. roeder (a) e Fg.5. Examples of“good” cheek detection. Table 2. Cheek line detection results Number Description (good, good) Both cheek lines are correctly located 204 (good, marginal One“good” - fit and one“ marginal"fit (marginal, marginal) pproximate locations are detected n both sides (good, wrong) 14 One cheek position is correct, the (marginal, wrong) 4 One cheek position is approximate the other is wrong Face contour extraction !73 g (c)(m,w) (d)(w,g) W (f) edge map F1g. 6. Examples of mixed cheek results a wrong location or completely missed. The edge in- edge detection results. In the case of Fig. 6(e), one cheek formation for the right cheek in(e)is not adcquate as is completely missed because there is no edge present shown in (0. In(a), the angles of the cheek lines are around the correct cheek location, as shown in part() inaccurate, although the positions are correct The Sometimes the cheek line detection can be thrown position of the left cheek line is off a few pixels in(b), off by the presence of glasses on the face, although it is and the position of one cheek line is actually along the successful in most cases right ear in(c). The lighting conditions were quite poor e If the subject is not facing straight ahead, only Then taking the picture in(d), where the left cheek line one cheek line can be detected consistently is completely missed. We expect the image quality to Sometimes the procedure detects the ear line Las be better than this case in realistic applications. in Fig. 6(c)] if that is much stronger than the cheek Table 2 shows the test results. Note that the cheek lines or when the subject is not facing straight ahead detection procedure located both cheeks with at least This situation can be avoided in most cases by a a marginal fit in 63 images. If a score of 2 is given to a properly chosen relevant subimage, but very skinny good"fit, 1 to a"marginal "fit, and 0 to a"wrong"faces do cause problems fit, our score is 270, which gives 270/(162 x 2)=83.3% of the maximum possible scoring 4. CHIN DETECTION The experiments were run on a Sparc IPC with a benchmark performance of 13.4 SPEC-marks 89. The Detection of chin lines involves searching for para erage total time for cheek detection is 3. 0s(after edge bolas in the edge image. Again the locations of the eyes detection). About 2.1s of this time were spent reading- and mouth are used to set the relevant subimage writing files The following limitations of the algorithm have been 4.1. The equation and parameters observed normally, a parabola is characterized by an equa- e 'This method, as are many other Hough-based tion with four or more parameters. Choosing an methods, is inherently dependent on the quality of equation with the lowest dimensionality reduces the 1174 X.li and n. roeder We now report our decisions on the issues men tioned in section 2 in the case of chin detection 1. Relevant subimage. A rectangular relevant sub xo =x-c(y-yo) image under the mouth is defined based on the posi tion of the eyes and mouth, which could be detected y the method given in Ref. 13 or that of Ref. 1. The position and the shape of the rectangle are determined by four variables, I, through 14, defined in Fig. 8. The actuali values(as a percentage of w ), observed from 28 test images, are recorded in Table 3. From our experi- Fig. 7. Parameter definitions in parabola detection ments on the remaining images, simply setting both l, and l2 to zero produces long enough chin lines, greatly simplifies the end point determination, and also re algorithmic complexity of the hT exponentially. (18) In duces the computational load a front-view face image most chins can be represented by an upright parabola, i. e. the central axis of the parabola is vertical. Thus, the following form is chosen 2. Parameter ranges. The ranges of xo and yo are defined according to the rectangle in Fig 8. The initial x-xo -c(-yo=0, range of c was set from -0.09 to-0.002. It has been observed that the minimum c value is.00223 and which is depicted in Fig. 7, where (xo, yo)is the vertex the maximum is -0.02973, which suggests that the of the parabola, and c controls how fast the parabola range for c could be restricted further. The target opens outward There are three ways to determine which accumula- precision of the c value is set to 0.0005, while the target tor cell should be incremented, knowing the coordi- precision for Xo and yo is a single pixel nates(x, y) of an edge pixel 3. Resolution of the accumulator. Recall equation (a) For each possible combination of (xo, yo)values, (2). The search area for c is L=0.088. Assuming b= 3 calculate the corresponding value for c and choosing m=6 will require k=5 iterations to reach the target precision g=0.0005. The value 6 fo mx and m, also gives approximately five iterations (y-y0)2 Therefore the search area sia)is divided into6x6x6 (b) For each possible combination of (xo, c) values, cells for all i calculate the corresponding value for yo 4. The use of edge strength and angle information yo=y士 The edge strength is used as described in Section 2. The angle information is discarded because of the complex ity of the computation needed to use it. (c) For each possible combination of (o, c)value alculate the corresponding value for xo 5. Neighborhood of a peak. a 3 x 3 neighborhood X-C surrounding the peak cell is used for zooming in, that b1=b2=b=3 2 Consider the rate of change in x when y stays the same ax ax (y-y)2 Cc(y-yo), w 2 In this experiment ax/ac is often much larger than the others, meaning very small changes in c make drastic changes to the shape of the parabola. There fore, method (a]will produce unreliable results because of the sensitivity of c. Mapping from xp and c to yo is inconvenient as it involves a square root operation Our algorithm uses method (c) with the middle value of each cell being used to represent that cell. Fig 8. Relevant subimages for chin detection Table 3. Recorded parameter values for the chin regions 1 l2 l3 max77.1%66.7% 97.6%(without beard), 111.3%(with beard mIn 200% Face contour extraction 1175 6. Number of peaks. In some faces, two parabolas 7. End points of the result curve. When the parabola e necessary to specify the chin line, such as in is being drawn on the image, the endpoints are found Figs 9 (d)and 10(b). No neighboring cells are ignored by looking from the extreme horizontal limits of the when searching for the second peak in the top-level image space and moving towards the calculated vertex accumulator due to the relatively small size of the of the parabola until a group of at least eight connected accumulator and the fact that some parabolas are pixels (in the edge image)are found. Then, the extreme often close together in the image(similar xo and yo point of this group of eight is used as the end point values). The second peak will be inspected if its vote The number 8 was obtained from experiment count is at least 93% of that of the first peak, i.e. v2/V1> chin for chin =0.93 which is significantly larger than 4.2. Experimental results and discussion the value of cheek. This chin was obtained fror periments on 72 images. Further tuning would be The chin detection procedure has been tested using helpful in correlation with the face identification pro- 72 face images in which some edge points are present cess. Before refining the second peak in the accumula- around the chin area. The results are categorized as tor array, a five-pixel swath on each side of the calcu-"good", "marginal"and"wrong "in a similar way to lated para bola is set to 0 in the edge strength array in that of cheek detection. Table 4 shows the test results order to avoid that parabola in the next iteration Note that the chin detection procedure successfully Table 4. Chin line detection results Good The chin line is located correctl Marginal 9 The parabola(s)fit(s)the chin approximately Wrong The parabola(s)is (are)wrong 廉 (d) Fig 9. Examples of"good"results from chin detection X. Li and N roedEr located the chin with at least a marginal fit in 67 images Some of the "marginal"results are shown in Fig 10 (93.1% of the cases) In case(a), the parabola should turn earlier at the Figure 9 gives examples of good fit. In case(), the bottom. The fact that the parabola is of the correct separation of the beard and the neck is detected which shape but is a few pixels too far to the right gave this can be considered as desirable. In case(d), a small instance a rating of"marginal". Case(b)is considered parabola at the top and a larger parabola at the bot- a marginal result because two parabolas should have tom jointly describe the lower portion of the face been detected for the chin to describe the face, but only Contour one is detected. The resulting parabola in (c) is too (a C a Fig. ll. Examples of"wrong"results from chin detection.

...展开详情
试读 13P 人脸识别 眼睛嘴 轮廓都解决
立即下载 低至0.43元/次 身份认证VIP会员低至7折
一个资源只可评论一次,评论内容不能少于5个字
您会向同学/朋友/同事推荐我们的CSDN下载吗?
谢谢参与!您的真实评价是我们改进的动力~
上传资源赚钱or赚积分
最新推荐
人脸识别 眼睛嘴 轮廓都解决 31积分/C币 立即下载
1/13
人脸识别 眼睛嘴 轮廓都解决第1页
人脸识别 眼睛嘴 轮廓都解决第2页
人脸识别 眼睛嘴 轮廓都解决第3页

试读结束, 可继续读1页

31积分/C币 立即下载 >