YOLO(You Only Look Once)\ub294 \uac1d\uccb4 \ud0d0\uc9c0 \ubd84\uc57c\uc5d0\uc11c \ub9e4\uc6b0 \ud601\uc2e0\uc801\uc778 \uc811\uadfc\ubc95\uc73c\ub85c, \uc5ec\ub7ec \uac1d\uccb4\ub97c \uc2e4\uc2dc\uac04\uc73c\ub85c \uc778\uc2dd\ud560 \uc218 \uc788\ub294 \ub6f0\uc5b4\ub09c \ub2a5\ub825\uc744 \uac00\uc9c0\uace0 \uc788\uc2b5\ub2c8\ub2e4. YOLO\ub294 \ud2b9\ud788 \uc5f0\uad6c\uc790\ub4e4\uacfc \uac1c\ubc1c\uc790\ub4e4 \uc0ac\uc774\uc5d0\uc11c \uc778\uae30\uac00 \uc788\uc73c\uba70, \uc2dc\uac04\uc5d0 \ub530\ub77c \uc5ec\ub7ec \ubc84\uc804\uc774 \uac1c\ubc1c\ub418\uc5c8\uc2b5\ub2c8\ub2e4. \uc774 \uae00\uc5d0\uc11c\ub294 YOLO\uc758 \ub2e4\uc591\ud55c \ubc84\uc804\uc778 YOLOv3, YOLOv4, YOLOv5\uc758 \ud2b9\uc9d5, \ucc28\uc774\uc810, \uadf8\ub9ac\uace0 \uc5b4\ub5a4 \uacbd\uc6b0\uc5d0 \uc5b4\ub5a4 \ubc84\uc804\uc744 \uc120\ud0dd\ud574\uc57c \ud558\ub294\uc9c0\ub97c \uc790\uc138\ud788 \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
YOLO\ub294 2016\ub144\uc5d0 \ucc98\uc74c \uc18c\uac1c\ub418\uc5c8\uc2b5\ub2c8\ub2e4. \uc774\uc804\uc758 \uac1d\uccb4 \ud0d0\uc9c0 \uc54c\uace0\ub9ac\uc998\ub4e4\uacfc \ub2ec\ub9ac, YOLO\ub294 \uc774\ubbf8\uc9c0 \uc804\uccb4\ub97c \ud55c \ubc88\uc5d0 \ucc98\ub9ac\ud558\uc5ec \uac1d\uccb4\ub97c \ud0d0\uc9c0\ud569\ub2c8\ub2e4. \uc774\ub294 \uc804\ud1b5\uc801\uc778 \ubc29\ubc95\ub4e4(\uc608: R-CNN \uacc4\uc5f4)\uc774 \uc9c0\uc5ed \uc81c\uc548(region proposal)\uc744 \uba3c\uc800 \uc0dd\uc131\ud55c \ud6c4, \ud574\ub2f9 \uc601\uc5ed\uc5d0 \ub300\ud574 \ubd84\ub958\uae30\ub97c \uc801\uc6a9\ud558\ub294 \ubc29\uc2dd\uacfc \ub2ec\ub9ac, \ud6e8\uc52c \ube60\ub974\uace0 \ud6a8\uc728\uc801\uc785\ub2c8\ub2e4. YOLO\ub294 \uc785\ub825 \uc774\ubbf8\uc9c0\uc5d0\uc11c \uac1d\uccb4\uc758 \uacbd\uacc4 \ubc15\uc2a4\uc640 \ud074\ub798\uc2a4 \ud655\ub960\uc744 \ub3d9\uc2dc\uc5d0 \uc608\uce21\ud558\ub294 \ub2e8\uc77c \uc2e0\uacbd\ub9dd\uc73c\ub85c \uad6c\uc131\ub429\ub2c8\ub2e4.<\/p>\n
YOLOv3\ub294 2018\ub144\uc5d0 \ubc1c\ud45c\ub41c \ubc84\uc804\uc73c\ub85c, \uc774\uc804 \ubc84\uc804\uc778 YOLOv2\uc5d0 \ube44\ud574 \uc5ec\ub7ec \uac00\uc9c0 \uac1c\uc120\uc774 \uc788\uc5c8\uc2b5\ub2c8\ub2e4. \uc8fc\uc694 \ud2b9\uc9d5\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
YOLOv3\uc758 \uc131\ub2a5\uc740 COCO \ub370\uc774\ud130\uc14b\uc5d0\uc11c mAP(mean Average Precision) 57.9%\ub85c \ud3c9\uac00\ub418\uc5c8\uc73c\uba70, \uac1d\uccb4 \ud0d0\uc9c0 \ubd84\uc57c\uc5d0\uc11c \uc5ec\uc804\ud788 \ub110\ub9ac \uc0ac\uc6a9\ub418\uace0 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
import cv2\nimport numpy as np\n\n# YOLOv3 \ubaa8\ub378 \ub85c\ub4dc\nnet = cv2.dnn.readNet(\"yolov3.weights\", \"yolov3.cfg\")\nlayer_names = net.getLayerNames()\noutput_layers = [layer_names[i - 1] for i in net.getUnconnectedOutLayers()]\n\n# \uc774\ubbf8\uc9c0 \ub85c\ub4dc\nimg = cv2.imread(\"image.jpg\")\nheight, width, channels = img.shape\n\n# \uc774\ubbf8\uc9c0 \uc804\ucc98\ub9ac\nblob = cv2.dnn.blobFromImage(img, 0.00392, (416, 416), (0, 0, 0), True, crop=False)\nnet.setInput(blob)\nouts = net.forward(output_layers)\n\n# \ud0d0\uc9c0\ub41c \uc815\ubcf4 \ucd9c\ub825\nclass_ids, confidences, boxes = [], [], []\nfor out in outs:\n for detection in out:\n scores = detection[5:]\n class_id = np.argmax(scores)\n confidence = scores[class_id]\n if confidence > 0.5:\n center_x = int(detection[0] * width)\n center_y = int(detection[1] * height)\n w = int(detection[2] * width)\n h = int(detection[3] * height)\n\n # \uacbd\uacc4 \ubc15\uc2a4 \uc88c\ud45c\n x = int(center_x - w \/ 2)\n y = int(center_y - h \/ 2)\n boxes.append([x, y, w, h])\n confidences.append(float(confidence))\n class_ids.append(class_id)\n\n# Non-Maximum Suppression\nindexes = cv2.dnn.NMSBoxes(boxes, confidences, 0.5, 0.4)\nfor i in indexes:\n i = i[0]\n box = boxes[i]\n x, y, w, h = box\n label = str(classes[class_ids[i]])\n cv2.rectangle(img, (x, y), (x + w, y + h), (0, 255, 0), 2)\n cv2.putText(img, label, (x, y + 30), cv2.FONT_HERSHEY_SIMPLEX, 3, (0, 255, 0), 3)\n\ncv2.imshow(\"Image\", img)\ncv2.waitKey(0)\ncv2.destroyAllWindows()<\/code><\/pre>\n3. YOLOv4<\/h2>\n
YOLOv4\ub294 2020\ub144\uc5d0 \ubc1c\ud45c\ub418\uc5c8\uc73c\uba70, YOLOv3\uc758 \uc131\ub2a5\uc744 \uac1c\uc120\ud558\uae30 \uc704\ud574 \uc5ec\ub7ec \ud601\uc2e0\uc744 \ub3c4\uc785\ud588\uc2b5\ub2c8\ub2e4. \uc8fc\uc694 \ud2b9\uc9d5\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
\n- \ub2e4\uc591\ud55c \ucd5c\uc801\ud654 \uae30\ubc95<\/strong>: \uc804\uc774 \ud559\uc2b5(transfer learning)\uacfc \ub370\uc774\ud130 \uc99d\uac15(data augmentation) \uae30\ubc95\uc744 \ud1b5\ud569\ud558\uc5ec \ubaa8\ub378\uc758 \ud6a8\uc728\uc131\uc744 \ub192\uc600\uc2b5\ub2c8\ub2e4.<\/li>\n
- CSPNet\uc758 \uc0ac\uc6a9<\/strong>: Cross Stage Partial Networks (CSPNet)\ub97c \ub3c4\uc785\ud558\uc5ec \ub354 \ub098\uc740 \ud2b9\uc9d5 \ud45c\ud604\uc744 \ud558\ub3c4\ub85d \ud588\uc2b5\ub2c8\ub2e4.<\/li>\n
- \ud55c\uce35 \uac1c\uc120\ub41c \uc131\ub2a5<\/strong>: YOLOv4\ub294 COCO \ub370\uc774\ud130\uc14b\uc5d0\uc11c 43.5 FPS\uc758 \uc18d\ub3c4\uc5d0\uc11c\ub3c4 mAP 43.5%\ub97c \ub2ec\uc131\ud558\uc5ec, \ub9ce\uc740 \uc2e4\uc81c \uc560\ud50c\ub9ac\ucf00\uc774\uc158\uc5d0\uc11c \uc4f0\uc77c \uc218 \uc788\ub294 \uac15\ub825\ud55c \uc194\ub8e8\uc158\uc774 \ub418\uc5c8\uc2b5\ub2c8\ub2e4.<\/li>\n<\/ul>\n
YOLOv4 \uc608\uc81c \ucf54\ub4dc<\/h3>\n# YOLOv4 \uc608\uc81c\ub294 YOLOv3\uc640 \ub9e4\uc6b0 \uc720\uc0ac\ud569\ub2c8\ub2e4.\n# YOLOv4\uc758 \uac00\uc911\uce58\uc640 \uad6c\uc131 \ud30c\uc77c\uc744 \uc774\uc6a9\ud558\uc5ec \uc544\ub798\uc640 \uac19\uc774 \ub85c\ub4dc\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.\n\nnet = cv2.dnn.readNet(\"yolov4.weights\", \"yolov4.cfg\")\n\n# \ub098\uba38\uc9c0 \ucf54\ub4dc\ub294 YOLOv3\uc640 \ub3d9\uc77c\ud569\ub2c8\ub2e4...\n# \uc774\ubbf8\uc9c0 \uc804\ucc98\ub9ac, \uc608\uce21, \uacb0\uacfc \uc2dc\uac01\ud654\ub294 \ub3d9\uc77c\ud558\uac8c \uc9c4\ud589\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/code><\/pre>\n4. YOLOv5<\/h2>\n
YOLOv5\ub294 2020\ub144\uc5d0 \ubc1c\ud45c\ub41c \ube44\uacf5\uc2dd\uc801\uc778 YOLO \ubc84\uc804\uc73c\ub85c, Ultralytics\uc5d0 \uc758\ud574 \uac1c\ubc1c\ub418\uc5c8\uc2b5\ub2c8\ub2e4. YOLOv5\ub294 YOLOv4\ubcf4\ub2e4 \ub354 \uac04\uacb0\ud558\uace0, \uc27d\uac8c \uc0ac\uc6a9\ud560 \uc218 \uc788\ub3c4\ub85d \ub9ce\uc740 \ucee4\ubba4\ub2c8\ud2f0\uc758 \uae30\uc5ec\ub97c \ubc1b\uc558\uc2b5\ub2c8\ub2e4. \uc8fc\uc694 \ud2b9\uc9d5\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
\n- PyTorch \uae30\ubc18 \uad6c\ud604<\/strong>: YOLOv5\ub294 PyTorch\ub85c \uad6c\ud604\ub418\uc5b4 \uc788\uc5b4, \ucf54\ub4dc\uac00 \uac04\uacb0\ud558\uace0 \uc0ac\uc6a9\uc790 \uce5c\ud654\uc801\uc785\ub2c8\ub2e4.<\/li>\n
- \ubaa8\ub378 \ud06c\uae30\uc640 \uc18d\ub3c4<\/strong>: YOLOv5\ub294 \ud06c\uae30\uc640 \uc18d\ub3c4 \uba74\uc5d0\uc11c \ub2e4\uc591\ud55c \uc635\uc158\uc744 \uc81c\uacf5\ud558\uc5ec, \uc0ac\uc6a9\uc790\uac00 \uc694\uad6c\uc5d0 \ub9de\uac8c \uc120\ud0dd\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/li>\n
- \ud5a5\uc0c1\ub41c \uc131\ub2a5<\/strong>: YOLOv5\ub294 COCO \ub370\uc774\ud130\uc14b\uc5d0\uc11c 50 FPS \uc774\uc0c1\uc758 \uc18d\ub3c4\ub85c mAP 50% \uc774\uc0c1\uc758 \uc131\ub2a5\uc744 \ubcf4\uc5ec\uc8fc\uba70 \uc0ac\uc6a9\ud558\uae30 \uc26c\uc6b4 \uc778\ud130\ud398\uc774\uc2a4\ub85c \ube60\ub974\uac8c \ud655\uc0b0\ub418\uc5c8\uc2b5\ub2c8\ub2e4.<\/li>\n<\/ul>\n
YOLOv5 \uc608\uc81c \ucf54\ub4dc<\/h3>\n!pip install torch torchvision # PyTorch \uc124\uce58\n\nfrom yolov5 import YOLOv5 # YOLOv5 \ub77c\uc774\ube0c\ub7ec\ub9ac \uc784\ud3ec\ud2b8\n\nmodel = YOLOv5(\"yolov5s.pt\") # \ubaa8\ub378 \ub85c\ub4dc\n\n# \uc774\ubbf8\uc9c0 \ub85c\ub4dc\nimg = \"image.jpg\"\nresults = model.predict(img)\n\n# \uacb0\uacfc \uc2dc\uac01\ud654\nresults.show()\nresults.save() # \uc800\uc7a5\ub41c \uc774\ubbf8\uc9c0 \ud655\uc778\uc6a9<\/code><\/pre>\n5. YOLO \ubc84\uc804 \uc120\ud0dd \uae30\uc900<\/h2>\n
\ub2e4\uc591\ud55c YOLO \ubc84\uc804\uc774 \uc874\uc7ac\ud558\uc9c0\ub9cc, \uc801\ud569\ud55c \ubc84\uc804\uc744 \uc120\ud0dd\ud558\uae30 \uc704\ud574\uc11c\ub294 \uba87 \uac00\uc9c0 \uc694\uc18c\ub97c \uace0\ub824\ud574\uc57c \ud569\ub2c8\ub2e4:<\/p>\n
\n- \uc18d\ub3c4<\/strong>: \uc2e4\uc2dc\uac04 \uc131\ub2a5\uc774 \uc911\uc694\ud55c \uc560\ud50c\ub9ac\ucf00\uc774\uc158\uc5d0\uc11c\ub294 YOLOv4\ub098 YOLOv5\uac00 \uc801\ud569\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/li>\n
- \uc815\ud655\ub3c4<\/strong>: \ub192\uc740 \uc815\ud655\ub3c4\uac00 \ud544\uc694\ud558\ub2e4\uba74 YOLOv4\uac00 \uc88b\uc740 \uc120\ud0dd\uc77c \uc218 \uc788\uc2b5\ub2c8\ub2e4. YOLOv5\ub3c4 \ucda9\ubd84\ud55c \uc815\ud655\ub3c4\ub97c \uc81c\uacf5\ud558\uc9c0\ub9cc, YOLOv4\uc640 \ube44\uad50\ud560 \ub54c \uc870\uae08 \ub0ae\uc744 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/li>\n
- \uac1c\ubc1c \ud3b8\uc758\uc131<\/strong>: YOLOv5\ub294 PyTorch\ub85c \uad6c\ud604\ub418\uc5b4 \uc788\uc5b4, \ube60\ub978 \uac1c\ubc1c\uacfc \ud29c\ub2dd\uc774 \uac00\ub2a5\ud569\ub2c8\ub2e4.<\/li>\n
- \ubaa8\ub378 \ud06c\uae30<\/strong>: \uacbd\ub7c9\ud654\ub41c \ubaa8\ub378\uc774 \ud544\uc694\ud558\ub2e4\uba74 \uc791\uc740 \ubc84\uc804\uc758 YOLOv5\ub97c \uace0\ub824\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/li>\n<\/ul>\n
\uacb0\ub860<\/h2>\n
YOLO \uacc4\uc5f4\uc740 \uac1d\uccb4 \ud0d0\uc9c0 \ubd84\uc57c\uc5d0\uc11c \ub9e4\uc6b0 \uc911\uc694\ud55c \uc5ed\ud560\uc744 \ud558\uace0 \uc788\uc73c\uba70, \uac01\uac01\uc758 \ubc84\uc804\uc774 \ubc1c\uc804\ud558\uba74\uc11c \ub354 \ub192\uc740 \uc131\ub2a5\uacfc \uc0ac\uc6a9 \ud3b8\uc758\uc131\uc744 \uc81c\uacf5\ud569\ub2c8\ub2e4. YOLOv3\ub294 \uc5ec\uc804\ud788 \ub9ce\uc740 \uc560\ud50c\ub9ac\ucf00\uc774\uc158\uc5d0\uc11c \uc0ac\uc6a9\ub418\uc9c0\ub9cc, YOLOv4\uc640 YOLOv5\uc758 \ucd9c\uc2dc\ub85c \uc778\ud574 \ub354 \ub098\uc740 \uc131\ub2a5\uc744 \uc81c\uacf5\ud558\ub294 \uc635\uc158\uc774 \ub9ce\uc544\uc84c\uc2b5\ub2c8\ub2e4. \uac1c\uc778\uc758 \ud544\uc694\uc5d0 \ub530\ub77c \uc801\uc808\ud55c YOLO \ubc84\uc804\uc744 \uc120\ud0dd\ud558\uc5ec \uc0ac\uc6a9\ud558\uc2dc\uae30 \ubc14\ub78d\ub2c8\ub2e4.<\/p>\n
<\/body><\/p>\n","protected":false},"excerpt":{"rendered":"