本文的文字及圖片來源於網絡,僅供學習、交流使用,不具有任何商業用途,版權歸原作者所有,如有問題請及時聯系我們以作處理。
作者： 機器學習與統計學
PS：如有需要Python學習資料的小伙伴可以加點擊下方鏈接自行獲取
http://note.youdao.com/noteshare?id=3054cce4add8a909e784ad934f956cef
隨着現代圖像處理和人工智能技術的快速發展，不少學者嘗試講CV應用到教學領域，能夠代替老師去閱卷，將老師從繁雜勞累的閱卷中解放出來，從而進一步有效的推動教學質量上一個台階。
傳統的人工閱卷，工作繁瑣，效率低下，進度難以控制且容易出現試卷遺漏未改、登分失誤等現象。
現代的“機器閱卷”，工作便捷、效率高、易操作，只需要一個相機（手機），拍照即可獲取成績，可以導入Excel表格便於存檔管理。
下面我們從代碼實現的角度來解釋一下我們這個簡易答題卡識別系統的工作原理。第一步，導入工具包及一系列的預處理
1 import numpy as np 2 import argparse 3 import imutils 4 import cv2 5 # 設置參數 6 ap = argparse.ArgumentParser() 7 ap.add_argument("-i", "--image", default="test_01.png") 8 args = vars(ap.parse_args()) 9 # 正確答案 10 ANSWER_KEY = {0: 1, 1: 4, 2: 0, 3: 3, 4: 1} # 11 def order_points(pts): 12 # 一共4個坐標點 13 rect = np.zeros((4, 2), dtype = "float32") 14 15 # 按順序找到對應坐標0,1,2,3分別是 左上，右上，右下，左下 16 # 計算左上，右下 17 s = pts.sum(axis = 1) 18 rect[0] = pts[np.argmin(s)] 19 rect[2] = pts[np.argmax(s)] 20 # 計算右上和左下 21 diff = np.diff(pts, axis = 1) 22 rect[1] = pts[np.argmin(diff)] 23 rect[3] = pts[np.argmax(diff)] 24 return rect 25 26 def four_point_transform(image, pts): 27 # 獲取輸入坐標點 28 rect = order_points(pts) 29 (tl, tr, br, bl) = rect 30 # 計算輸入的w和h值 31 widthA = np.sqrt(((br[0]-bl[0])** 2) + ((br[1]-bl[1])**2)) 32 widthB = np.sqrt(((tr[0] -tl[0]) ** 2) + ((tr[1] - tl[1]) ** 2)) 33 maxWidth = max(int(widthA), int(widthB)) 34 heightA = np.sqrt(((tr[0]-br[0])**2)+((tr[1]-br[1])**2)) 35 heightB = np.sqrt(((tl[0]-bl[0])**2)+((tl[1]-bl[1])**2)) 36 maxHeight = max(int(heightA), int(heightB)) 37 # 變換后對應坐標位置 38 dst = np.array([ 39 [0, 0], 40 [maxWidth - 1, 0], 41 [maxWidth - 1, maxHeight - 1], 42 [0, maxHeight - 1]], dtype = "float32") 43 # 計算變換矩陣 44 M = cv2.getPerspectiveTransform(rect, dst) 45 warped = cv2.warpPerspective(image, M, (maxWidth, maxHeight)) 46 return warped # 返回變換后結果 47 48 def sort_contours(cnts, method="left-to-right"): 49 reverse = False 50 i = 0 51 if method == "right-to-left" or method == "bottom-to-top": 52 reverse = True 53 if method == "top-to-bottom" or method == "bottom-to-top": 54 i = 1 55 boundingBoxes = [cv2.boundingRect(c) for c in cnts] 56 (cnts, boundingBoxes) = zip(*sorted(zip(cnts, boundingBoxes), 57 key=lambda b: b[1][i], reverse=reverse)) 58 return cnts, boundingBoxes 59 def cv_show(name,img): 60 cv2.imshow(name, img) 61 cv2.waitKey(0) 62 cv2.destroyAllWindows() 63 64 65 image = cv2.imread(args["image"]) 66 contours_img = image.copy() 67 gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY) 68 blurred = cv2.GaussianBlur(gray, (5, 5), 0) 69 edged = cv2.Canny(blurred, 75, 200) 70 # 輪廓檢測 71 cnts = cv2.findContours(edged.copy(), cv2.RETR_EXTERNAL, 72 cv2.CHAIN_APPROX_SIMPLE)[1] 73 cv2.drawContours(contours_img,cnts,-1,(0,0,255),3) 74 docCnt = None 75 76 # 確保檢測到了 77 if len(cnts) > 0: 78 # 根據輪廓大小進行排序 79 cnts = sorted(cnts, key=cv2.contourArea, reverse=True) 80 for c in cnts: # 遍歷每一個輪廓 81 # 近似 82 peri = cv2.arcLength(c, True) 83 approx = cv2.approxPolyDP(c, 0.02 * peri, True) 84 # 准備做透視變換 85 if len(approx) == 4: 86 docCnt = approx 87 break 88 # 執行透視變換 89 warped = four_point_transform(gray, docCnt.reshape(4, 2)) 90 91 thresh = cv2.threshold(warped, 0, 255, 92 cv2.THRESH_BINARY_INV | cv2.THRESH_OTSU)[1] 93 thresh_Contours = thresh.copy() 94 # 找到每一個圓圈輪廓 95 cnts = cv2.findContours(thresh.copy(), cv2.RETR_EXTERNAL, 96 cv2.CHAIN_APPROX_SIMPLE)[1] 97 cv2.drawContours(thresh_Contours,cnts,-1,(0,0,255),3) 98 questionCnts = [] 99 for c in cnts:# 遍歷 100 # 計算比例和大小 101 (x, y, w, h) = cv2.boundingRect(c) 102 ar = w / float(h) 103 # 根據實際情況指定標准 104 if w >= 20 and h >= 20 and ar >= 0.9 and ar <= 1.1: 105 questionCnts.append(c) 106 # 按照從上到下進行排序 107 questionCnts = sort_contours(questionCnts, 108 method="top-to-bottom")[0] 109 correct = 0 110 # 每排有5個選項 111 for (q, i) in enumerate(np.arange(0, len(questionCnts), 5)): 112 cnts = sort_contours(questionCnts[i:i + 5])[0] 113 bubbled = None 114 for (j, c) in enumerate(cnts): # 遍歷每一個結果 115 # 使用mask來判斷結果 116 mask = np.zeros(thresh.shape, dtype="uint8") 117 cv2.drawContours(mask, [c], -1, 255, -1) #-1表示填充 118 # 通過計算非零點數量來算是否選擇這個答案 119 mask = cv2.bitwise_and(thresh, thresh, mask=mask) 120 total = cv2.countNonZero(mask) 121 # 通過閾值判斷 122 if bubbled is None or total > bubbled[0]: 123 bubbled = (total, j) 124 # 第二步，與正確答案進行對比 125 color = (0, 0, 255) 126 k = ANSWER_KEY[q] 127 # 判斷正確 128 if k == bubbled[1]: 129 color = (0, 255, 0) 130 correct += 1 131 cv2.drawContours(warped, [cnts[k]], -1, color, 3) #繪圖 132 133 #正確率的文本顯示 134 score = (correct / 5.0) * 100 135 print("[INFO] score: {:.2f}%".format(score)) 136 cv2.putText(warped, "{:.2f}%".format(score), (10, 30), 137 cv2.FONT_HERSHEY_SIMPLEX, 0.9, (0, 0, 255), 2) 138 cv2.imshow("Input", image) 139 cv2.imshow("Output", warped) 140 cv2.waitKey(0)
最終實現的效果如下：
.