|
opencv_1">利用opencv实现车牌检测
整体流程涉及5个部分
- 图像通道转换
- 对比度增强
- 边缘连接
- 二值化
- 边界区域裁剪
图像通道转换
将RGB图像转换为HSV图像,仅保留V通道。V通道表示颜色的明暗,常用于图像对比度拉伸、直方图均衡化等流程。
原图像:
V通道图像:
对比度增强
通过顶帽变换来实现对比度增强。顶帽变换用于提取图像的小区域和局部细节。白顶帽变换用于提取图像中比周围环境亮的小物体或细节;黑顶帽变换用于提取图像中比周围环境暗的小物体或细节。
白顶帽变换:
黑顶帽变化:
通过白顶帽、黑顶帽的联合处理: I e n h a n c e d = I o r i g i n a l + I w h i t e − t o p _ h a t − I b l a c k − t o p _ h a t I_{enhanced}=I_{original}+I_{white-top\_hat}-I_{black-top\_hat} Ienhanced=Ioriginal+Iwhite−top_hat−Iblack−top_hat,其中 I o r i g i n a l I_{original} Ioriginal表示原图像, I w h i t e − t o p _ h a t I_{white-top\_hat} Iwhite−top_hat表示白顶帽处理后图像, I b l a c k − t o p _ h a t I_{black-top\_hat} Iblack−top_hat表示黑顶帽处理后图像,得到对比度增强后的图像:
边缘连接
增强对比度后,很多车牌边缘不连续,例如
需要通过膨胀操作(Dilation Operation)来扩展边缘,实现边缘连接的目的。
添加膨胀操作后,图像转变为:
二值化
将单通道V图像转换为二值图像,具体策略为Adaptive thresholding
边界区域裁剪
- 首先,利用cv2.findContours检测边界,并且获得边界的层级(hierarchy)。
- 车牌检测可以理解为找到内边界,而整个图像的背景可以理解为是外边界。下图是检测出的内边界
对内边界进行阈值判断处理,过滤掉明显错误的情况。例如过滤面积小于2000的内边界(具体数值需要按照实际情况来定) - 对于每个内边界,计算外接最小的矩形(可以通过统计边界内最左、最上、最右、最下的点来合成矩形),作为初步检测框
- 有一些检测框可能包括多个车牌,宽度、高度比较大。对于这种情况,需要对检测框按照宽度、高度均匀分割。以下是一个高度过大的例子,需按高度均分
- 有一些车牌因为自身比较模糊,导致检测框不准确,可以通过统计信息来过滤掉,本方法暂不处理。例如
最终,整张图有41个车牌,通过上述方法,检测到了40个车牌,效果不错。漏检的车牌本身边缘不清晰,检测难度较大
消融实验
方法 | 最终图像检测框 | 车牌检测数量 |
---|---|---|
最终方法 | 40 | |
去掉对比度增强 | 39 | |
去掉边缘连接 | 39 | |
内边界面积过滤阈值4000 | 38 | |
内边界面积过滤阈值5000 | 38 |
代码
"""
主要的步骤为:
1)提取单通道图片,选项为 (灰度图片/HSV中的value分支)
2)提升对比度,选项为 (形态学中的顶帽/灰度拉伸)
3)边缘连接(膨胀)
4)二值化
5)利用findcontours函数找到边缘
6)裁剪图片,车牌图片存储
7) 对车牌预处理
8)方向矫正
9)车牌精确区域搜索
10) 字符分割
11) 字符识别
"""import cv2
import copy
import numpy as np
import math
import osdef SingleChannel(img) :"""用于车牌检测得到单通道图片,主要测试两种方式,灰度通道以及hsv中的v通道:param img: 输入图片:return:"""hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)hue, saturation, value = cv2.split(hsv)cv2.imshow("SingleChannel", value)return valuedef Contrast(img) :"""用于车牌检测利用tophat,提高图片对比度,:param img: 输入图片:return:"""kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (3, 3))# applying topHat/blackHat operationstopHat = cv2.morphologyEx(img, cv2.MORPH_TOPHAT, kernel)cv2.imshow("tophat", topHat)blackHat = cv2.morphologyEx(img, cv2.MORPH_BLACKHAT, kernel)cv2.imshow("blackhat", blackHat)add = cv2.add(img, topHat)subtract = cv2.subtract(add, blackHat)cv2.imshow('Constrast', subtract)return subtractdef threshold(img) :"""用于车牌检测采用cv2.adaptiveThreshold方法,对图片二值化:param img: 输入图像:return:"""thresh = cv2.adaptiveThreshold(img, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY_INV, 19, 9)cv2.imshow("thresh", thresh)return threshglobal crop_num
crop_num = 0def drawCoutrous(img_temp) :"""对输入图像查找内边缘,设置阈值,去除一些面积较小的内边缘:param img_temp: 输入图像,经过预处理:return:"""threshline = 2000imgCopy = copy.deepcopy(img_temp)contours, hierarchy = cv2.findContours(imgCopy, cv2.RETR_CCOMP, cv2.CHAIN_APPROX_SIMPLE)# print(len(contours), contours[0].shape)# print(hierarchy.shape)maxarea = 0conid = 0img_zero = np.zeros(img.shape)# print("img_zero.shape is : ",img_zero.shape)num_contours = 0contoursList = []for i in range(len(contours)) :if hierarchy[0][i][3] >= 0 :temparea = math.fabs(cv2.contourArea(contours[i]))# print(math.fabs(cv2.contourArea(contours[i])))if temparea > maxarea :conid = imaxarea = tempareaif temparea > threshline :num_contours += 1if num_contours % 7 == 0 :cv2.drawContours(img_zero, contours, i, (0,0,255),1)if num_contours % 7 == 1 :cv2.drawContours(img_zero, contours, i, (255,0,0),1)if num_contours % 7 == 2 :cv2.drawContours(img_zero, contours, i, (0,255,0),1)if num_contours % 7 == 3 :cv2.drawContours(img_zero, contours, i, (0,255,255),1)if num_contours % 7 == 4 :cv2.drawContours(img_zero, contours, i, (255,0,255),1)if num_contours % 7 == 5 :cv2.drawContours(img_zero, contours, i, (255,255,0),1)if num_contours % 7 == 6:cv2.drawContours(img_zero, contours, i, (255, 255, 255), 1)# print(contours[i].shape)contoursList.append(contours[i])# print("maxarea: ",maxarea)# print("number of contours is ", num_contours)# cv2.drawContours(img_zero, contours, conid, (0, 0, 255), 1)cv2.imshow("with contours",img_zero)return contoursListdef DrawRectangle(img, img_temp, ConList) :"""得到车牌边缘的的x,y坐标最小最大值,再原图上绘制bounding box,得到裁剪后的车牌图像:param img: 原图:param img_temp: 二值图像:param ConList: 图像的边缘轮廓:return: null"""length = len(ConList)rectanglePoint = np.zeros((length, 4, 1, 2), dtype = np.int32)img_zeros = np.zeros(img_temp.shape)img_copy = copy.deepcopy(img)img_copy_1 = copy.deepcopy(img)# print("img_zeros, length; ", img_zeros.shape, length)for i in range(length) :contours = ConList[i]minx, maxx, miny, maxy = 1e6, 0, 1e6, 0for index_num in range(contours.shape[0]) :if contours[index_num][0][0] < minx :minx = contours[index_num][0][0]if contours[index_num][0][0] > maxx :maxx = contours[index_num][0][0]if contours[index_num][0][1] < miny :miny = contours[index_num][0][1]if contours[index_num][0][1] > maxy :maxy = contours[index_num][0][1]# print(minx, maxx, miny, maxy)rectanglePoint[i][0][0][0], rectanglePoint[i][0][0][1] = minx, minyrectanglePoint[i][1][0][0], rectanglePoint[i][1][0][1] = minx, maxyrectanglePoint[i][2][0][0], rectanglePoint[i][2][0][1] = maxx, maxyrectanglePoint[i][3][0][0], rectanglePoint[i][3][0][1] = maxx, miny# rectanglePoint.dtype = np.int32# print(rectanglePoint[i].shape)crop_save(minx, maxx, miny, maxy, img_copy_1)# print("dx: ",maxx-minx,"dy: ",maxy-miny, "area: ", (maxx-minx)*(maxy-miny))cv2.polylines(img_copy, [rectanglePoint[i]], True, (0,0,255),2)cv2.imshow("img_zeros_haha", img_copy)def crop_save(minx, maxx, miny, maxy, img_original) :"""裁剪原图,根据minx,maxx,miny,maxy:param minx: x坐标最小值:param maxx: x坐标最大值:param miny: y坐标最小值:param maxy: y坐标最大值:param img_original: 由于需要将绘制结果再原图中显示,输入原图:return:"""global crop_numepsx = 60epsy = 30dx = maxx - minxdy = maxy - minyif dx == dy :returnif dx >= 600 - epsx :dx1, dx2, dx3, dx4 = minx, minx + 1 * int(dx / 3), minx + 2 * int(dx / 3), maxxsave_pth = './crop40/cropimg_' + str(crop_num) + '.jpg'# cv2.imwrite(save_pth, img_original[dx1:dx2, miny:maxy,:])cv2.imwrite(save_pth, img_original[miny:maxy, dx1:dx2, :])crop_num += 1save_pth = './crop40/cropimg_' + str(crop_num) + '.jpg'cv2.imwrite(save_pth, img_original[miny:maxy, dx2:dx3, :])crop_num += 1save_pth = './crop40/cropimg_' + str(crop_num) + '.jpg'cv2.imwrite(save_pth, img_original[miny:maxy, dx3:dx4, :])crop_num += 1elif dx >= 400 - epsx :dx1, dx2, dx3 = minx, minx + 1 * int(dx / 2), maxxsave_pth = './crop40/cropimg_' + str(crop_num) + '.jpg'cv2.imwrite(save_pth, img_original[miny:maxy, dx1:dx2, :])crop_num += 1save_pth = './crop40/cropimg_' + str(crop_num) + '.jpg'cv2.imwrite(save_pth, img_original[miny:maxy, dx2:dx3, :])crop_num += 1elif dy >= 240 - epsy :dy1, dy2, dy3, dy4 = miny, miny + 1 * int(dy / 3), miny + 2 * int(dy / 3), maxysave_pth = './crop40/cropimg_' + str(crop_num) + '.jpg'cv2.imwrite(save_pth, img_original[dy1: dy2, minx:maxx, :])crop_num += 1save_pth = './crop40/cropimg_' + str(crop_num) + '.jpg'cv2.imwrite(save_pth, img_original[dy2: dy3, minx:maxx, :])crop_num += 1save_pth = './crop40/cropimg_' + str(crop_num) + '.jpg'cv2.imwrite(save_pth, img_original[dy3: dy4, minx:maxx, :])crop_num += 1elif dy >= 160 - epsy :dy1, dy2, dy3 = miny, miny + 1 * int(dy / 2), maxysave_pth = './crop40/cropimg_' + str(crop_num) + '.jpg'cv2.imwrite(save_pth, img_original[dy1: dy2, minx:maxx, :])crop_num += 1save_pth = './crop40/cropimg_' + str(crop_num) + '.jpg'cv2.imwrite(save_pth, img_original[dy2: dy3, minx:maxx, :])crop_num += 1elif dx <= 200 + epsx :dx1, dx2 = minx, maxxsave_pth = './crop40/cropimg_' + str(crop_num) + '.jpg'cv2.imwrite(save_pth, img_original[miny:maxy, dx1:dx2, :])crop_num += 1else :passif __name__ == '__main__' :pth = 'License_plates.jpg'img = cv2.imread(pth)img = cv2.resize(img, (292 * 4, 173 * 4))cv2.imshow("original",img)# 1)提取单通道图片,选项为 (灰度图片/HSV中的value分支)singlechannel_img = SingleChannel(img)# 2)提升对比度contrast_img = Contrast(singlechannel_img)# contrast_img = singlechannel_img# 3)边缘连接(膨胀)kernel = np.ones((2, 2), np.uint8)dilation_img = cv2.dilate(contrast_img, kernel, iterations=1)cv2.imshow("dilate", dilation_img)# dilation_img = contrast_img# 4) 二值化threshold_img = threshold(dilation_img)# 5)利用findcontours函数找到边缘contoursList = drawCoutrous(threshold_img)# 6) 裁剪图片,车牌图片存储DrawRectangle(img, threshold_img, contoursList)cv2.waitKey()cv2.destroyAllWindows()