import numpy as np import cv2 import math from myGpuFeatures import myGpuFeatures # from scipy.stats import mode class Method(): outputAddress = "result/" isEvaluate = False evaluateFile = "evaluate.txt" isPrintLog = True featureMethod = "sift" # "sift","surf" or "orb" roiRatio = 0.1 # roi length for stitching in first direction searchRatio = 0.75 # 0.75 is common value for matches isGPUAvailable = False surfHessianThreshold = 100.0 surfNOctaves = 4 surfNOctaveLayers = 3 surfIsExtended = True surfKeypointsRatio = 0.01 surfIsUpright = False orbNfeatures = 5000 orbScaleFactor = 1.2 orbNlevels = 8 orbEdgeThreshold = 31 orbFirstLevel = 0 orbWTA_K = 2 orbPatchSize = 31 orbFastThreshold = 20 orbBlurForDescriptor = False orbMaxDistance = 30 offsetCaculate = "mode" # "mode" or "ransac" offsetEvaluate = 3 # 40 menas nums of matches for mode, 3.0 menas of matches for ransac isEnhance = False isClahe = False clipLimit = 20 tileSize = 5 def printAndWrite(self, content): """ :param content: :return: """ if self.isPrintLog: print(content) if self.isEvaluate: f = open(self.outputAddress + self.evaluateFile, "a") f.write(content) f.write("\n") f.close() def getROIRegionForIncreMethod(self, image, direction=1, order="first", searchRatio=0.1): """ :param image: :param direction: :param order: ‘first’or'second' :param searchRatio: :return: """ row, col = image.shape[:2] roiRegion = np.zeros(image.shape, np.uint8) if direction == 1: searchLength = np.floor(row * searchRatio).astype(int) if order == "first": roiRegion = image[row - searchLength:row, :] elif order == "second": roiRegion = image[0: searchLength, :] elif direction == 2: searchLength = np.floor(col * searchRatio).astype(int) if order == "first": roiRegion = image[:, col - searchLength:col] elif order == "second": roiRegion = image[:, 0: searchLength] elif direction == 3: searchLength = np.floor(row * searchRatio).astype(int) if order == "first": roiRegion = image[0: searchLength, :] elif order == "second": roiRegion = image[row - searchLength:row, :] elif direction == 4: searchLength = np.floor(col * searchRatio).astype(int) if order == "first": roiRegion = image[:, 0: searchLength] elif order == "second": roiRegion = image[:, col - searchLength:col] return roiRegion def getROIRegion(self, image, direction="horizontal", order="first", searchLength=150, searchLengthForLarge=-1): ''' :param originalImage: :param direction: :param order: :param searchLength: :param searchLengthForLarge: :return: :type searchLength: np.int ''' row, col = image.shape[:2] if direction == "horizontal" or direction == 2: if order == "first": if searchLengthForLarge == -1: roiRegion = image[:, col - searchLength:col] elif searchLengthForLarge > 0: roiRegion = image[row - searchLengthForLarge:row, col - searchLength:col] elif order == "second": if searchLengthForLarge == -1: roiRegion = image[:, 0: searchLength] elif searchLengthForLarge > 0: roiRegion = image[0:searchLengthForLarge, 0: searchLength] elif direction == "vertical" or direction == 1: if order == "first": if searchLengthForLarge == -1: roiRegion = image[row - searchLength:row, :] elif searchLengthForLarge > 0: roiRegion = image[row - searchLength:row, col - searchLengthForLarge:col] elif order == "second": if searchLengthForLarge == -1: roiRegion = image[0: searchLength, :] elif searchLengthForLarge > 0: roiRegion = image[0: searchLength, 0:searchLengthForLarge] return roiRegion def getOffsetByMode(self, kpsA, kpsB, matches, offsetEvaluate = 10): """ :param kpsA: :param kpsB: :param matches: :param offsetEvaluate: :return:(totalStatus, [dx, dy]), totalStatus ,[dx, dy][0, 0] """ totalStatus = True if len(matches) == 0: totalStatus = False return (totalStatus, [0, 0]) dxList = []; dyList = []; for trainIdx, queryIdx in matches: ptA = (kpsA[queryIdx][1], kpsA[queryIdx][0]) ptB = (kpsB[trainIdx][1], kpsB[trainIdx][0]) # dxList.append(int(round(ptA[0] - ptB[0]))) # dyList.append(int(round(ptA[1] - ptB[1]))) if int(ptA[0] - ptB[0]) == 0 and int(ptA[1] - ptB[1]) == 0: continue dxList.append(int(ptA[0] - ptB[0])) dyList.append(int(ptA[1] - ptB[1])) if len(dxList) == 0: dxList.append(0); dyList.append(0) # Get Mode offset in [dxList, dyList], thanks for clovermini zipped = zip(dxList, dyList) zip_list = list(zipped) zip_dict = dict((a, zip_list.count(a)) for a in zip_list) zip_dict_sorted = dict(sorted(zip_dict.items(), key=lambda x: x[1], reverse=True)) dx = list(zip_dict_sorted)[0][0] dy = list(zip_dict_sorted)[0][1] num = zip_dict_sorted[list(zip_dict_sorted)[0]] # print("dx = " + str(dx) + ", dy = " + str(dy) + ", num = " + str(num)) if num < offsetEvaluate: totalStatus = False # self.printAndWrite(" In Mode, The number of num is " + str(num) + " and the number of offsetEvaluate is "+str(offsetEvaluate)) return (totalStatus, [dx, dy]) def getOffsetByRansac(self, kpsA, kpsB, matches, offsetEvaluate=100): """ :param kpsA: :param kpsB: :param matches: :param offsetEvaluate: :return:(totalStatus, [dx, dy]), totalStatus ,[dx, dy][0, 0] """ totalStatus = False ptsA = np.float32([kpsA[i] for (_, i) in matches]) ptsB = np.float32([kpsB[i] for (i, _) in matches]) if len(matches) == 0: return (totalStatus, [0, 0], 0) H1 = cv2.getAffineTransform(ptsA, ptsB) # print("H1") # print(H1) (H, status) = cv2.findHomography(ptsA, ptsB, cv2.RANSAC, 3, 0.9) trueCount = 0 for i in range(0, len(status)): if status[i] == True: trueCount = trueCount + 1 if trueCount >= offsetEvaluate: totalStatus = True adjustH = H.copy() adjustH[0, 2] = 0;adjustH[1, 2] = 0 adjustH[2, 0] = 0;adjustH[2, 1] = 0 return (totalStatus ,[np.round(np.array(H).astype(np.int)[1,2]) * (-1), np.round(np.array(H).astype(np.int)[0,2]) * (-1)], adjustH) else: return (totalStatus, [0, 0], 0) def npToListForKeypoints(self, array): ''' :Convert array to List, used for keypoints from GPUDLL to python List :param array: array from GPUDLL :return: ''' kps = [] row, col = array.shape for i in range(row): kps.append([array[i, 0], array[i, 1]]) return kps def npToListForMatches(self, array): ''' :Convert array to List, used for DMatches from GPUDLL to python List :param array: array from GPUDLL :return: ''' descritpors = [] row, col = array.shape for i in range(row): descritpors.append((array[i, 0], array[i, 1])) return descritpors def npToKpsAndDescriptors(self, array): """ :param array: :return: """ kps = [] descriptors = array[:, :, 1] for i in range(array.shape[0]): kps.append([array[i, 0, 0], array[i, 1, 0]]) return (kps, descriptors) def detectAndDescribe(self, image, featureMethod): ''' :param image: :return:(kps, features) ''' descriptor = cv2.SIFT_create() if self.isGPUAvailable == False: # CPU mode if featureMethod == "sift": descriptor = cv2.SIFT_create() elif featureMethod == "surf": #descriptor = cv2.SURF_create() pass elif featureMethod == "orb": descriptor = cv2.ORB_create(self.orbNfeatures, self.orbScaleFactor, self.orbNlevels, self.orbEdgeThreshold, self.orbFirstLevel, self.orbWTA_K, 0, self.orbPatchSize, self.orbFastThreshold) kps, features = descriptor.detectAndCompute(image, None) kps = np.float32([kp.pt for kp in kps]) else: # GPU mode if featureMethod == "sift": descriptor = cv2.SIFT_create() kps, features = descriptor.detectAndCompute(image, None) kps = np.float32([kp.pt for kp in kps]) elif featureMethod == "surf": kps, features = self.npToKpsAndDescriptors(myGpuFeatures.detectAndDescribeBySurf(image, self.surfHessianThreshold, self.surfNOctaves,self.surfNOctaveLayers, self.surfIsExtended, self.surfKeypointsRatio, self.surfIsUpright)) elif featureMethod == "orb": kps, features = self.npToKpsAndDescriptors(myGpuFeatures.detectAndDescribeByOrb(image, self.orbNfeatures, self.orbScaleFactor, self.orbNlevels, self.orbEdgeThreshold, self.orbFirstLevel, self.orbWTA_K, 0, self.orbPatchSize, self.orbFastThreshold, self.orbBlurForDescriptor)) return (kps, features) def matchDescriptors(self, featuresA, featuresB): ''' :param featuresA: :param featuresB: :return:matches ''' if self.isGPUAvailable == False: # CPU Mode if self.featureMethod == "surf" or self.featureMethod == "sift": matcher = cv2.DescriptorMatcher_create("BruteForce") rawMatches = matcher.knnMatch(featuresA, featuresB, 2) matches = [] for m in rawMatches: if len(m) == 2 and m[0].distance < m[1].distance * self.searchRatio: matches.append((m[0].trainIdx, m[0].queryIdx)) elif self.featureMethod == "orb": matcher = cv2.DescriptorMatcher_create("BruteForce-Hamming") rawMatches = matcher.match(featuresA, featuresB) matches = [] for m in rawMatches: matches.append((m.trainIdx, m.queryIdx)) # self.printAndWrite(" The number of matches is " + str(len(matches))) else: # GPU Mode if self.featureMethod == "surf": matches = self.npToListForMatches(myGpuFeatures.matchDescriptors(np.array(featuresA), np.array(featuresB), 2, self.searchRatio)) elif self.featureMethod == "orb": matches = self.npToListForMatches(myGpuFeatures.matchDescriptors(np.array(featuresA), np.array(featuresB), 3, self.orbMaxDistance)) return matches def resizeImg(self, image, resizeTimes, interMethod = cv2.INTER_AREA): """ :param image: :param resizeTimes: :param interMethod: :return: """ (h, w) = image.shape resizeH = int(h * resizeTimes) resizeW = int(w * resizeTimes) # cv2.INTER_AREA return cv2.resize(image, (resizeW, resizeH), interpolation=interMethod) def rectifyFinalImg(self, image, regionLength = 10): """ :param image: :param regionLength: :return: """ (h, w) = image.shape print("h:" + str(h)) print("w:" + str(w)) upperLeft = np.sum(image[0: regionLength, 0: regionLength]) upperRight = np.sum(image[0: regionLength, w - regionLength: w]) bottomLeft = np.sum(image[h - regionLength: h, 0: regionLength]) bottomRight = np.sum(image[h - regionLength: h, w - regionLength: w]) zeroCol = image[:, 0] noneZeroNum = np.count_nonzero(zeroCol) zeroNum = h - noneZeroNum print("noneZeroNum:" + str(noneZeroNum)) print("zeroNum:" + str(zeroNum)) print("division:" + str(noneZeroNum / h)) if (noneZeroNum / h) < 0.3: resultImage = image elif upperLeft == 0 and bottomRight == 0 and upperRight != 0 and bottomLeft != 0: print(1) center = (w // 2, h // 2) print(w) print(h) angle = math.atan(center[1] / center[0] * 180 / math.pi) print(str(angle)) M = cv2.getRotationMatrix2D(center, -1 * angle, 1.0) print(M) resultImage = cv2.warpAffine(image, M, (w, h)) elif upperLeft != 0 and bottomRight != 0 and upperRight == 0 and bottomLeft == 0: print(2) center = (w // 2, h // 2) angle = math.atan(center[1] / center[0] * 180 / math.pi) print(str(angle)) M = cv2.getRotationMatrix2D(center, angle, 1.0) resultImage = cv2.warpAffine(image, M, (w, h)) else: resultImage = image return resultImage