序:想要真正准确的的自动标注,的确不太现实,都能准确的自动标注了,还训练模型干嘛!
所以本文所写方法是小量数据集预训练模型后,自动打标最后微调。
(上图是我的文件夹格式,将自己预训练后的模型放到指定位置)
代码包含调用yolo模型。废话不多说!
# coding=utf-8 ''' author : Helen date : 2020-11-12 16:15 ''' import cv2 import numpy as np import os import xml.dom.minidom # import pypinyin import time def ModelYoloV3(frame, confidence=0.5, threshold=0.4): # 加载文件路径 yolo_dir = './previousTrainModel' # YOLO文件路径 weightsPath = os.path.join(yolo_dir, 'myVLPChar_yolov3_21000.weights') # 权重文件 configPath = os.path.join(yolo_dir, 'myVLPChar_yolov3.cfg') # 配置文件 labelsPath = os.path.join(yolo_dir, 'myVLPCharData.names') # label名称 print("[INFO] loading YOLO from disk...") # # 可以打印下信息 # 加载网络、配置权重 net = cv2.dnn.readNetFromDarknet(configPath, weightsPath) # # 利用下载的文件 # net.setPreferableBackend(cv.dnn.DNN_BACKEND_OPENCV) # net.setPreferableTarget(cv.dnn.DNN_TARGET_CPU) CONFIDENCE = confidence # 过滤弱检测的最小概率,默认0.5 THRESHOLD = threshold # 非最大值抑制阈值默认0.4 # 加载图片、转为blob格式、送入网络输入层 img = frame.copy() blobImg = cv2.dnn.blobFromImage(img, 1.0 / 255.0, (416, 416), None, True, False) # # net需要的输入是blob格式的,用blobFromImage这个函数来转格式 net.setInput(blobImg) # # 调用setInput函数将图片送入输入层 # 获取网络输出层信息(所有输出层的名字),设定并前向传播 outInfo = net.getUnconnectedOutLayersNames() # # 前面的yolov3架构也讲了,yolo在每个scale都有输出,outInfo是每个scale的名字信息,供net.forward使用 start = time.time() layerOutputs = net.forward(outInfo) # 得到各个输出层的、各个检测框等信息,是二维结构。 end = time.time() print("[INFO] YOLO took {:.6f} seconds".format(end - start)) # # 可以打印下信息 # 拿到图片尺寸 (H, W) = img.shape[:2] # 过滤layerOutputs # layerOutputs的第1维的元素内容: [center_x, center_y, width, height, objectness, N-class score data] # 过滤后的结果放入: boxes = [] # 所有边界框(各层结果放一起) confidences = [] # 所有置信度,概率 classIDs = [] # 所有分类ID rectsAndClasses = [] # 所有概率超过阈值的框位置列表 # # 1)过滤掉置信度低的框框 for out in layerOutputs: # 各个输出层 for detection in out: # 各个框框 # 拿到置信度 scores = detection[5:] # 各个类别的置信度 classID = np.argmax(scores) # 最高置信度的id即为分类id confidence = scores[classID] # 拿到置信度 # 根据置信度筛查 if confidence > CONFIDENCE: box = detection[0:4] * np.array([W, H, W, H]) # 将边界框放会图片尺寸 (centerX, centerY, width, height) = box.astype("int") x = int(centerX - (width / 2)) y = int(centerY - (height / 2)) boxes.append([x, y, int(width), int(height)]) confidences.append(float(confidence)) classIDs.append(classID) # # 2)应用非最大值抑制(non-maxima suppression,nms)进一步筛掉 idxs = cv2.dnn.NMSBoxes(boxes, confidences, CONFIDENCE, THRESHOLD) # boxes中,保留的box的索引index存入idxs # 得到labels列表 with open(labelsPath, 'rt') as f: labels = f.read().rstrip(' ').split(' ') # 应用检测结果 np.random.seed(42) COLORS = np.random.randint(0, 255, size=(len(labels), 3), dtype="uint8") # 框框显示颜色,每一类有不同的颜色,每种颜色都是由RGB三个值组成的,所以size为(len(labels), 3) if len(idxs) > 0: for i in idxs.flatten(): # indxs是二维的,第0维是输出层,所以这里把它展平成1维 (x, y) = (boxes[i][0], boxes[i][1]) (w, h) = (boxes[i][2], boxes[i][3]) color = [int(c) for c in COLORS[classIDs[i]]] cv2.rectangle(img, (x, y), (x + w, y + h), color, 3) # 线条粗细为2px text = "{}: {:.4f}".format(labels[int(classIDs[i])], confidences[i]) cv2.putText(img, text, (x, y - 5), cv2.FONT_HERSHEY_SIMPLEX, 0.5, color, 2) # cv.FONT_HERSHEY_SIMPLEX字体风格、0.5字体大小、粗细2px print(labels[int(classIDs[i])], ":", confidences[i]) rectsAndClasses.append([x, y, w, h, labels[int(classIDs[i])], confidences[i], color]) return img, rectsAndClasses # 返回画过框的图片和NMS后的框列表 def genXML(imgName, shape, labels, classes, xmlPath): # new_txtname = imgName.split('.')[0] # # 创建空的Dom文档对象 doc = xml.dom.minidom.Document() # 创建根结点,根节点名为 annotation annotation = doc.createElement('annotation') # 根节点 # 将根节点添加到Dom文档对象中 doc.appendChild(annotation) # folder节点 folder = doc.createElement('folder') # 创建一个名叫folder的节点 # 内容写入 folder_text = doc.createTextNode('JPEGImages') # folder节点里面要写的内容 folder.appendChild(folder_text) # 添加到folder节点下,如果是内容,节点内容createTextNode类型,就作为内容写入;如果是createElement类型,就作为子节点添加进去 annotation.appendChild(folder) # 之后将添加好内容的folder节点,作为子节点添加到annotation节点中 # filename节点 filename = doc.createElement('filename') filename_text = doc.createTextNode(str(new_txtname) + '.jpg') filename.appendChild(filename_text) # annotation.appendChild(filename) # path节点 path = doc.createElement('path') path_text = doc.createTextNode('E:\darknet-master\build\darknet\myVLPCharData\JPEGImages\%s.jpg' % new_txtname) # 框架路径,根据自己修改 path.appendChild(path_text) # annotation.appendChild(path) # sourch节点 source = doc.createElement('source') # database = doc.createElement('database') database_text = doc.createTextNode('Unknown') database.appendChild(database_text) # source.appendChild(database) # annotation.appendChild(source) # size节点 size = doc.createElement('size') width = doc.createElement('width') width_text = doc.createTextNode(str(shape[1])) width.appendChild(width_text) size.appendChild(width) height = doc.createElement('height') height_text = doc.createTextNode(str(shape[0])) height.appendChild(height_text) size.appendChild(height) depth = doc.createElement('depth') depth_text = doc.createTextNode(str(shape[-1])) depth.appendChild(depth_text) size.appendChild(depth) # annotation.appendChild(size) # segmented节点 segmented = doc.createElement('segmented') segmented_text = doc.createTextNode('0') segmented.appendChild(segmented_text) # annotation.appendChild(segmented) # object节点 for [y1, y2, x1, x2], pChar in zip(labels, classes): # 分类坐标和分类名称 object = doc.createElement('object') name = doc.createElement('name') name_text = doc.createTextNode(pChar) # 这个地方是标签的name,也就是分类名称 name.appendChild(name_text) object.appendChild(name) pose = doc.createElement('pose') pose_text = doc.createTextNode("Unspecified") pose.appendChild(pose_text) object.appendChild(pose) truncated = doc.createElement('truncated') truncated_text = doc.createTextNode("0") truncated.appendChild(truncated_text) object.appendChild(truncated) difficult = doc.createElement('difficult') difficult_text = doc.createTextNode("0") difficult.appendChild(difficult_text) object.appendChild(difficult) bndbox = doc.createElement('bndbox') # xmin = doc.createElement('xmin') xmin_text = doc.createTextNode(str(x1)) xmin.appendChild(xmin_text) bndbox.appendChild(xmin) # ymin = doc.createElement('ymin') ymin_text = doc.createTextNode(str(y1)) ymin.appendChild(ymin_text) bndbox.appendChild(ymin) # xmax = doc.createElement('xmax') xmax_text = doc.createTextNode(str(x2)) xmax.appendChild(xmax_text) bndbox.appendChild(xmax) # ymax = doc.createElement('ymax') ymax_text = doc.createTextNode(str(y2)) ymax.appendChild(ymax_text) bndbox.appendChild(ymax) # object.appendChild(bndbox) # annotation.appendChild(object) # 写入xml文本文件中 if not os.path.exists(xmlPath): os.mkdir(xmlPath) fp = open(xmlPath + '/%s.xml' % new_txtname, 'w+') doc.writexml(fp, indent=' ', addindent=' ', newl='', encoding='utf-8') fp.close() def run(imgPath): filesName = os.listdir(imgPath) # 读取所有图片名 for f in filesName: p = os.path.join(imgPath, f) # 单独图片的路径 frame = cv2.imread(p) classesImg, rectsAndClasses = ModelYoloV3(frame) cv2.imshow("Labels", classesImg) labels = [] classes = [] for x, y, w, h, cn, confidence, color in rectsAndClasses: X1 = x Y1 = y X2 = x + w Y2 = y + h labels.append([Y1, Y2, X1, X2]) # 分类坐标 classes.append(cn) # 分类名称 genXML(f, frame.shape, labels, classes, 'E:\darknet-master\build\xml') # args: 图片名,图片shape, 标签坐标列表, 类名列表, xml文件保存路径 cv2.waitKey(10) class AutoLabels: def __init__(self): pass if __name__ == '__main__': path = 'E:\darknet-master\build\image' run(path)