\uc544\ub798\uc758 \uc774\ubbf8\uc9c0\ub294 \uc22b\uc790 0\uc5d0 \ub300\ud55c \uc785\ub825 \uc774\ubbf8\uc9c0(\uc67c\ucabd)\ub97c deskew \ud568\uc218\uc5d0 \uc801\uc6a9\ud55c \uacb0\uacfc(\uc624\ub978\ucabd)\uc785\ub2c8\ub2e4.<\/p>\n
![\"\"](\"http:\/\/www.gisdeveloper.co.kr\/wp-content\/uploads\/2019\/05\/deskew.jpg\")
<\/p>\n
\ub2e4\uc74c\uc740 \uac01 \uc140\uc758 HOG \ub514\uc2a4\ud06c\ub9bd\ud130\ub97c \ucc3e\ub294 \uac83\uc785\ub2c8\ub2e4. \uc774\ub97c \uc704\ud574, X\ucd95\uacfc Y\ucd95 \ubc29\ud5a5\uc73c\ub85c \uac01 \uc140\uc758 Sobel \uacb0\uacfc\ub97c \uc5bb\uc2b5\ub2c8\ub2e4. \uadf8\ub9ac\uace0\ub294 \uc774 \uacb0\uacfc\uc5d0 \ub300\ud55c \uac01 \ud53d\uc140\uc5d0\uc11c \uadf8\ub808\ub514\uc5b8\ud2b8(Gradient)\uc758 \ud06c\uae30\uc640 \ubc29\ud5a5\uc744 \uc5bb\uc2b5\ub2c8\ub2e4. \uc774 \uadf8\ub808\ub514\uc5b8\ud2b8\ub294 16\uac1c\uc758 \uc815\uc218\uac12\uc73c\ub85c \ubcc0\ud658\ud569\ub2c8\ub2e4. \uc774 \uc774\ubbf8\uc9c0\ub97c 4\uac1c\uc758 \ubd80\ubd84 \uc0ac\uac01\ud615\uc73c\ub85c \ubd84\ud560\ud569\ub2c8\ub2e4. \uac01\uac01\uc758 \ubd80\ubd84 \uc0ac\uac01\ud615\uc5d0 \ub300\ud574, \uc774\ub4e4\uc758 \ud06c\uae30\uac12\uc744 \uac00\uc9c0\ub294 \uac00\uc911\uce58 \ubc29\ud5a5(16 Bins)\uc758 \ud788\uc2a4\ud1a0\uadf8\ub7a8\uc744 \uacc4\uc0b0\ud569\ub2c8\ub2e4. \uadf8\ub7ec\uba74 \uac01\uac01\uc758 \ubd80\ubd84 \uc0ac\uac01\ud615\uc744 \ud1b5\ud574 16\uac1c\uc758 \uac12\uc744 \uac00\uc9c4 \ubca1\ud130\uac00 \uc0dd\uc131\ub429\ub2c8\ub2e4. 4\uac1c\uc758 \uac01 \ubca1\ud130(4\uac1c\uc758 \ubd80\ubd84 \uc0ac\uac01\ud615)\ub97c \ud569\uce58\uba74 64\uac1c\uc758 \uac12\uc744 \uac16\ub294 \ud2b9\uc9d5 \ubca1\ud130\ub97c \uc5bb\uac8c \ub429\ub2c8\ub2e4. \uc774 \ud2b9\uc9d5 \ubca1\ud130\ub97c \ud6c8\ub828 \ub370\uc774\ud130\ub85c \uc0ac\uc6a9\ud569\ub2c8\ub2e4. \uc544\ub798\uc758 \ucf54\ub4dc\ub294 \uc9c0\uae08\uae4c\uc9c0\uc758 \uc124\uba85\uc5d0 \ub300\ud55c \ucf54\ub4dc\uc785\ub2c8\ub2e4.<\/p>\n
\r\nbin_n = 16 # Number of bins\r\n\r\ndef hog(img):\r\n gx = cv2.Sobel(img, cv2.CV_32F, 1, 0)\r\n gy = cv2.Sobel(img, cv2.CV_32F, 0, 1)\r\n mag, ang = cv2.cartToPolar(gx, gy)\r\n bins = np.int32(bin_n*ang\/(2*np.pi)) # quantizing binvalues in (0...16)\r\n bin_cells = bins[:10,:10], bins[10:,:10], bins[:10,10:], bins[10:,10:]\r\n mag_cells = mag[:10,:10], mag[10:,:10], mag[:10,10:], mag[10:,10:]\r\n hists = [np.bincount(b.ravel(), m.ravel(), bin_n) for b, m in zip(bin_cells, mag_cells)]\r\n hist = np.hstack(hists) # hist is a 64 bit vector\r\n return hist\r\n<\/pre>\n\ub9c8\uc9c0\ub9c9\uc73c\ub85c, \uc774\uc804 \uacbd\uc6b0\uc5d0\uc11c\ucc98\ub7fc, \uc22b\uc790\uac00 \uc4f0\uc5ec\uc9c4 \ud070 \uc774\ubbf8\uc9c0\ub97c \uac01 \uc140\ub85c \uc870\uac01\ub0c5\ub2c8\ub2e4. \uac01 \ubb38\uc790\uc5d0 \ub300\ud574, 250\uac1c\uc758 \uc140\uc740 \ud6c8\ub828 \ub370\uc774\ud130\ub85c.. \ub098\uba38\uc9c0\ub294 \uc2dc\ud5d8 \ud14c\uc774\ud130\ub85c \uc0ac\uc6a9\ud560 \uac83\uc785\ub2c8\ub2e4. \uc544\ub798\ub294 \uc774\uc5d0 \ub300\ud55c \uc124\uba85\uacfc \uc55e\uc11c \uc5b8\uae09\ud55c 2\uac1c\uc758 \ud568\uc218 \ubaa8\ub450\ub97c \ud569\ud55c \uc804\uccb4 \ucf54\ub4dc\uc785\ub2c8\ub2e4.<\/p>\n
\r\nimport cv2\r\nimport numpy as np\r\n\r\nSZ=20\r\n\r\naffine_flags = cv2.WARP_INVERSE_MAP|cv2.INTER_LINEAR\r\n\r\ndef deskew(img):\r\n m = cv2.moments(img)\r\n if abs(m['mu02']) < 1e-2:\r\n return img.copy()\r\n skew = m['mu11']\/m['mu02']\r\n M = np.float32([[1, skew, -0.5*SZ*skew], [0, 1, 0]])\r\n img = cv2.warpAffine(img,M,(SZ, SZ),flags=affine_flags)\r\n return img\r\n\r\nbin_n = 16 # Number of bins\r\n\r\ndef hog(img):\r\n gx = cv2.Sobel(img, cv2.CV_32F, 1, 0)\r\n gy = cv2.Sobel(img, cv2.CV_32F, 0, 1)\r\n mag, ang = cv2.cartToPolar(gx, gy)\r\n bins = np.int32(bin_n*ang\/(2*np.pi)) # quantizing binvalues in (0...16)\r\n bin_cells = bins[:10,:10], bins[10:,:10], bins[:10,10:], bins[10:,10:]\r\n mag_cells = mag[:10,:10], mag[10:,:10], mag[:10,10:], mag[10:,10:]\r\n hists = [np.bincount(b.ravel(), m.ravel(), bin_n) for b, m in zip(bin_cells, mag_cells)]\r\n hist = np.hstack(hists) # hist is a 64 bit vector\r\n return hist\r\n\r\nimg = cv2.imread('.\/data\/digits.png',0)\r\n\r\ncells = [np.hsplit(row,100) for row in np.vsplit(img,50)]\r\n\r\n# First half is trainData, remaining is testData\r\ntrain_cells = [ i[:50] for i in cells ]\r\ntest_cells = [ i[50:] for i in cells]\r\n\r\n###### Now training ########################\r\ndeskewed = [list(map(deskew,row)) for row in train_cells]\r\nhogdata = [list(map(hog,row)) for row in deskewed]\r\ntrainData = np.float32(hogdata).reshape(-1,64)\r\nresponses = np.repeat(np.arange(10),250)[:,np.newaxis]\r\n\r\nsvm = cv2.ml.SVM_create()\r\nsvm.setKernel(cv2.ml.SVM_LINEAR)\r\nsvm.setType(cv2.ml.SVM_C_SVC)\r\nsvm.setC(2.67)\r\nsvm.setGamma(5.383)\r\n\r\nsvm.train(trainData, cv2.ml.ROW_SAMPLE, responses)\r\nsvm.save('svm_data.dat')\r\n\r\n###### Now testing ########################\r\ndeskewed = [list(map(deskew,row)) for row in test_cells]\r\nhogdata = [list(map(hog,row)) for row in deskewed]\r\ntestData = np.float32(hogdata).reshape(-1,bin_n*4)\r\nresult = svm.predict(testData)[1]\r\n\r\n####### Check Accuracy ########################\r\nmask = result==responses\r\ncorrect = np.count_nonzero(mask)\r\nprint(correct*100.0\/result.size)\r\n<\/pre>\n\uacb0\uacfc\ub294 93.8\ub85c\uc368 \uc778\uc2dd \uc131\uacf5\ub960\uc785\ub2c8\ub2e4. kNN\uc5d0\uc11c\uc758 \uc778\uc2dd\ub960\uc740 91.76\uc600\ub294\ub370.. \uadf8\ubcf4\ub2e4 \ub354 \ub098\uc740 \uc778\uc2dd\ub960\uc744 \ubcf4\uc774\uace0 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n","protected":false},"excerpt":{"rendered":"
\uc774 \uae00\uc758 \uc6d0\ubb38\uc740 https:\/\/opencv-python-tutroals.readthedocs.io\/en\/latest\/py_tutorials\/py_ml\/py_svm\/py_svm_opencv\/py_svm_opencv.html \uc785\ub2c8\ub2e4. \uc774 \uae00\uc740 kNN \ub300\uc2e0 SVM\ub97c \uc0ac\uc6a9\ud55c OCR \uae30\ub2a5\uc744 OpenCV\ub85c \uad6c\ud604\ud558\ub294 \uc608\uc81c\uc5d0 \ub300\ud55c \uc124\uba85\uc785\ub2c8\ub2e4. kNN\uc5d0\uc11c\ub294 \ud2b9\uc9d5 \ubca1\ud130(Feature Vector)\ub85c\uc368 \ud53d\uc140\uc758 \ubc1d\uae30(Intensity)\ub97c \uc0ac\uc6a9\ud588\uc2b5\ub2c8\ub2e4. \uc774\ubc88\uc5d0\ub294 \ud2b9\uc9d5 \ubca1\ud130\ub85c\uc368 HOF(Histogram of Oriented Gradients)\ub97c \uc0ac\uc6a9\ud558\ub3c4\ub85d \ud558\uaca0\uc2b5\ub2c8\ub2e4. HOG\ub97c \ucc3e\uae30 \uc804\uc5d0, \uba3c\uc800 \uc774\ubbf8\uc9c0\uc758 Moments(\uc815\ud655\ud788\ub294 Second Order Moments)\ub97c \uc0ac\uc6a9\ud574 \uae30\uc6b8\uc5b4\uc9c4 \uc774\ubbf8\uc9c0\ub97c \ubcf4\uc815\ud569\ub2c8\ub2e4. \uba3c\uc800 deskew() \ud568\uc218\ub97c \uc815\uc758\ud560\ud150\ub370, \uc774 \ud568\uc218\ub294 \uc22b\uc790 \uc774\ubbf8\uc9c0\ub97c … <\/p>\n