人工智能-深度学习-计算机视觉-LBP级联+CNN 回归定位车牌

#coding=utf-8 import numpy as np import cv2 import os import time import training_judging as tj; from keras.models import Sequential from keras.layers import Dense, Dropout, Activation, Flatten from keras.layers import Convolution2D, MaxPooling2D from keras.optimizers import SGD import keras import numpy as np watch_cascade = cv2.CascadeClassifier('./out-long/cascade.xml') count = 0; model = tj.constructmodel((13,55,1)) model.load_weights("./judge1.h5") extend_scale = 1; shape = [140,180] def getTransformMatrix(origin_shape,shape_= (140,180)): pts1 = np.float32([[0,0],[origin_shape[1],0],[origin_shape[1],origin_shape[0]],[0,origin_shape[0]]]) pts2 = np.float32([[0,0],[shape_[1],0],[shape_[1],shape_[0]],[0,shape_[0]] ]) M = cv2.getPerspectiveTransform(pts1,pts2) return M def resize_with_points(pts,shape,image): M = getTransformMatrix(image.shape,shape) points_ = np.dot(M, np.float32(np.vstack((pts.T, np.array([1, 1, 1, 1]).T))))[0:2, 0:4].T.astype(np.int) return points_ def foutpoint(p_points): p_points = p_points[0] xp = np.array([[p_points[0]*shape[1],p_points[1]*shape[0]], [p_points[2] * shape[1], p_points[3] * shape[0]], [p_points[4] * shape[1], p_points[5] * shape[0]], [p_points[6] * shape[1], p_points[7] * shape[0]] ]) return xp def drawPoints(img,p_points): # img = img*255 # img = img.astype(np.uint8); # p = p_points.reshape([4,2]); # r = r_points.reshape([4, 2]); # print p p_points = p_points[0] # print p_points p = np.array([[p_points[0]*shape[1],p_points[1]*shape[0]], [p_points[2] * shape[1], p_points[3] * shape[0]], [p_points[4] * shape[1], p_points[5] * shape[0]], [p_points[6] * shape[1], p_points[7] * shape[0]] ]) # print p # r = np.array([[r_points[0]*200,r_points[1]*72], # [r_points[2] * 200, r_points[3] * 72], # [r_points[4] * 200, r_points[5] * 72], # [r_points[6] * 200, r_points[7] * 72] # ]) # points[0][0] = p_points[0] 200, points[0][1] / 72, # points[1][0] / 200, points[1][1] / 72, # points[2][0] / 200, points[2][1] / 72, # points[3][0] / 200, points[3][1] / 72] for one in p: cv2.circle(img,(int(one[0]),int(one[1])),1,(0,255,0),2) # print "R",r # for one in r: # cv2.circle(img,(int(one[0]),int(one[1])),3,(0,255,0)) cv2.imshow("img", img) cv2.waitKey(0) def rectSrceen(pts,rect): zeroadd = np.array(rect[:2]) return np.array(pts) + zeroadd def norm(X): #X = cv2.resize(X,(28,28)) P = cv2.cvtColor(X,cv2.COLOR_RGB2GRAY) P = P.astype(np.float32)/255 P = np.expand_dims(P,2) return P def getmodel(path): model = Sequential(); model.add(Convolution2D(1, 7, 7, border_mode='valid', input_shape=(140, 180, 1), subsample=(2, 2))) model.add(Activation('relu')) model.add(MaxPooling2D(pool_size=(3, 3))) model.add(Convolution2D(1, 3, 3, border_mode='valid', subsample=(2, 2))) model.add(Convolution2D(32, 5, 5, border_mode='valid', subsample=(2, 2))) model.add(MaxPooling2D(pool_size=(2, 2))) model.add(Activation('relu')) # model.add(Dropout(0.1)) model.add(Flatten()) model.add(Dense(128, init='normal')) model.add(Activation('relu')) model.add(Dense(8, init='normal')) model.add(Activation('relu')) model.load_weights(path) return model def packFilename(origin_filename,position,type=".png"): name,suffix = os.path.splitext(origin_filename) filename = name+"_" for coor in position: filename+=str(coor)+"_" filename+=type; return filename models = getmodel("model_t5_regression.h5") path = "./general_test" for parent,dirnames,filenames in os.walk("/Users/yujinke/Desktop/EasyPR/resources/image/general_test"): for filename in filenames: # while(1): # filename = filenames[np.random.randint(0,len(filenames))] path = os.path.join(parent,filename) if path.endswith(".jpg"): img = cv2.imread(path) img_bak = cv2.imread(path,cv2.IMREAD_GRAYSCALE ) img_bak = cv2.GaussianBlur(img_bak,(3,3),2) gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) t0 = time.time() watches = watch_cascade.detectMultiScale(gray, 1.1, 2,minSize=(76 , 18)) print time.time() - t0 for (x, y, w, h) in watches: #cv2.rectangle(img, (x, y), (x + w, y + h), (255, 0, 0), 1) patch = img_bak[y:y + h, x:x + w] #cv2.imwrite("./cascade_results/"+str(count)+".png",patch) # patch[:,:,0] = cv2.equalizeHist(patch[:,:,0]) # patch[:, :, 1] = cv2.equalizeHist(patch[:, :, 1]) # patch[:, :, 2] = cv2.equalizeHist(patch[:, :, 2]) vector = tj.norm(patch) prob = model.predict(np.array([vector]))[0] print prob prob_1 = prob[0] prob_2 = prob[1] #cv2.putText(img,str(round(prob_1,2)),(x,y+15),cv2.FONT_HERSHEY_COMPLEX,0.5,(255,0,255)) # if prob_2<0.7 or prob_1>0.5: # cv2.rectangle(img, (x, y), (x + w, y + h), (255, 255, 0), 1) extend_w = int(w*extend_scale) extend_h = int(h*extend_scale) start_corr_y = y - int(extend_h*1.0) end_coor_y =y + extend_h + int(extend_h*1.0) start_corr_x = x - int(extend_w*0.4) end_coor_x = x + extend_w + int(extend_w*0.4) if start_corr_x < 0 : start_corr_x = 0; if start_corr_y < 0 : start_corr_y = 0; if end_coor_x>img.shape[1]: end_coor_x = img.shape[1] if end_coor_y>img.shape[0]: end_coor_y = img.shape[0] extend_patch = img[start_corr_y:end_coor_y, start_corr_x:end_coor_x] cv2.imshow("extend_patch",extend_patch) cv2.waitKey(0) extend_patch_resize = cv2.resize(extend_patch, (180, 140)) image = norm(extend_patch_resize) vector = models.predict(np.array([image])) points = foutpoint(vector) points = resize_with_points(points,extend_patch.shape,extend_patch_resize) points = rectSrceen(points,[start_corr_x,start_corr_y]) remap_points = np.array([[0,0],[136,0],[136,36],[0,36]],dtype=np.float32); print points,remap_points M = cv2.getPerspectiveTransform(np.array(points,dtype=np.float32),remap_points) warp_image = cv2.warpPerspective(img,M,(136,36)) cv2.imshow("warp_image",warp_image) for one in points: cv2.circle(img, (int(one[0]), int(one[1])), 1, (0, 255, 0), 2) #drawPoints(extend_patch, vector) count += 1; print count cv2.imshow("img",img) cv2.waitKey(0)