\n\ubcf8 \uac15\uc88c\ub294 OpenCV\ub97c \uc774\uc6a9\ud558\uc5ec \uba38\uc2e0\ub7ec\ub2dd \ubaa8\ub378\uc744 \ud6a8\uc728\uc801\uc73c\ub85c \ud65c\uc6a9\ud558\ub294 \ubc29\ubc95\uc744 \ub2e4\ub8f9\ub2c8\ub2e4. \ud2b9\ud788, ONNX(Open Neural Network Exchange) \ud615\uc2dd\uc73c\ub85c \ubaa8\ub378\uc744 \ubcc0\ud658\ud558\ub294 \uacfc\uc815\uacfc \uadf8 \ubaa8\ub378\uc744 OpenCV\uc5d0\uc11c \uc2e4\ud589\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc124\uba85\ud558\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n<\/header>\n\n1. ONNX\ub780 \ubb34\uc5c7\uc778\uac00?<\/h2>\n
ONNX\ub294 \ub2e4\uc591\ud55c \uba38\uc2e0\ub7ec\ub2dd \ud504\ub808\uc784\uc6cc\ud06c \uac04\uc758 \uc0c1\ud638 \uc6b4\uc6a9\uc131\uc744 \ub192\uc774\uae30 \uc704\ud574 \uc124\uacc4\ub41c \uac1c\ubc29\ud615 \ud3ec\ub9f7\uc785\ub2c8\ub2e4. ONNX \ud615\uc2dd\uc744 \uc0ac\uc6a9\ud558\uba74 TensorFlow, PyTorch \ub4f1 \uc5ec\ub7ec \ud504\ub808\uc784\uc6cc\ud06c\uc5d0\uc11c \ud559\uc2b5\ud55c \ubaa8\ub378\uc744 \uc190\uc27d\uac8c \ubcc0\ud658\ud558\uace0 \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. ONNX\ub294 \ub2e4\uc74c\uacfc \uac19\uc740 \uc7a5\uc810\uc744 \uc81c\uacf5\ud569\ub2c8\ub2e4:<\/p>\n
\n- \ubaa8\ub378 \ud638\ud658\uc131: \ub2e4\uc591\ud55c \ud50c\ub7ab\ud3fc\uacfc \uc5b8\uc5b4\uc5d0\uc11c \uc27d\uac8c \ub85c\ub4dc\ud558\ub294 \uac83\uc774 \uac00\ub2a5\ud558\ub2e4.<\/li>\n
- \ud37c\ud3ec\uba3c\uc2a4: \ucd5c\uc801\ud654\ub41c \uc5f0\uc0b0\uc744 \ud1b5\ud574 \ubaa8\ub378\uc758 \uc2e4\ud589\uc18d\ub3c4\ub97c \ud5a5\uc0c1\uc2dc\ud0ac \uc218 \uc788\ub2e4.<\/li>\n
- \uc720\uc5f0\uc131: \ub2e4\uc911 \ud504\ub808\uc784\uc6cc\ud06c \uc9c0\uc6d0\uc73c\ub85c \uc778\ud574 \ud2b9\uc815 \ud504\ub808\uc784\uc6cc\ud06c\uc5d0 \uad6d\ud55c\ub418\uc9c0 \uc54a\ub294\ub2e4.<\/li>\n<\/ul>\n<\/section>\n\n
2. OpenCV\ub780 \ubb34\uc5c7\uc778\uac00?<\/h2>\n
OpenCV(Open Source Computer Vision Library)\ub294 \ucef4\ud4e8\ud130 \ube44\uc804 \ubc0f \uba38\uc2e0\ub7ec\ub2dd\uc744 \uc704\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac\ub85c\uc11c, C++, Python, Java \ub4f1 \ub2e4\uc591\ud55c \uc5b8\uc5b4\ub97c \uc9c0\uc6d0\ud569\ub2c8\ub2e4. OpenCV\ub294 \uc774\ubbf8\uc9c0 \ucc98\ub9ac, \ube44\ub514\uc624 \ubd84\uc11d, \ube44\ub514\uc624 \uac10\uc2dc, \ub85c\ubd07 \ube44\uc804 \ub4f1\uc758 \ubd84\uc57c\uc5d0\uc11c \uad11\ubc94\uc704\ud558\uac8c \uc0ac\uc6a9\ub429\ub2c8\ub2e4.<\/p>\n<\/section>\n\n3. ONNX \ubaa8\ub378 \ubcc0\ud658 \uacfc\uc815<\/h2>\n3.1. PyTorch \ubaa8\ub378\uc744 ONNX\ub85c \ubcc0\ud658\ud558\uae30<\/h3>\n
PyTorch\uc5d0\uc11c \ud6c8\ub828\ud55c \ubaa8\ub378\uc744 ONNX \ud615\uc2dd\uc73c\ub85c \ubcc0\ud658\ud558\ub294 \ubc29\ubc95\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, \uac04\ub2e8\ud55c \uc774\ubbf8\uc9c0 \ubd84\ub958 \ubaa8\ub378\uc744 \ub9cc\ub4e4\uc5b4 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
\nimport torch\nimport torch.onnx\nimport torchvision.models as models\n\n# \uae30\ubcf8 ResNet \ubaa8\ub378\uc744 \uc0dd\uc131\nmodel = models.resnet18(pretrained=True)\nmodel.eval()\n\n# \ub354\ubbf8 \uc785\ub825 \ub370\uc774\ud130 \uc0dd\uc131 (\ubc30\uce58 \ud06c\uae30 1, \ucc44\ub110 3, \ub192\uc774 224, \ub108\ube44 224)\ndummy_input = torch.randn(1, 3, 224, 224)\n\n# ONNX \ud615\uc2dd\uc73c\ub85c \ubaa8\ub378 \uc800\uc7a5\ntorch.onnx.export(model, dummy_input, \"resnet18.onnx\", \n verbose=True, \n input_names=['input'], \n output_names=['output'])\n <\/code><\/pre>\n\uc704 \ucf54\ub4dc\uc5d0\uc11c\ub294 PyTorch\uc5d0\uc11c \uc81c\uacf5\ud558\ub294 \ubbf8\ub9ac \ud6c8\ub828\ub41c ResNet-18 \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud558\uc5ec .onnx \ud30c\uc77c\ub85c \ubcc0\ud658\ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4. torch.onnx.export<\/code> \ud568\uc218\uc758 \ub9e4\uac1c\ubcc0\uc218\ub85c \ubaa8\ub378, \ub354\ubbf8 \uc785\ub825 \ub370\uc774\ud130, \uc800\uc7a5\ud560 \ud30c\uc77c \uc774\ub984\uc744 \uc9c0\uc815\ud569\ub2c8\ub2e4. \ub610\ud55c \uc785\ub825 \ubc0f \ucd9c\ub825 \ud150\uc11c\uc758 \uc774\ub984\uc744 \uc815\uc758\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n3.2. TensorFlow \ubaa8\ub378\uc744 ONNX\ub85c \ubcc0\ud658\ud558\uae30<\/h3>\n
TensorFlow\uc5d0\uc11c \ud6c8\ub828\ub41c \ubaa8\ub378\ub3c4 ONNX \ud615\uc2dd\uc73c\ub85c \ubcc0\ud658\uc774 \uac00\ub2a5\ud569\ub2c8\ub2e4. TensorFlow\uc5d0\uc11c ONNX\ub85c \ubcc0\ud658\ud558\ub824\uba74 tf2onnx<\/code> \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ub2e4\uc74c\uc740 TensorFlow \ubaa8\ub378\uc744 ONNX \ud615\uc2dd\uc73c\ub85c \ubcc0\ud658\ud558\ub294 \uc608\uc81c\uc785\ub2c8\ub2e4.<\/p>\n\nimport tensorflow as tf\nimport tf2onnx\n\n# \uc608\uc81c \ubaa8\ub378 \uc0dd\uc131\nmodel = tf.keras.applications.MobileNetV2(weights='imagenet', input_shape=(224, 224, 3))\n\n# \uc785\ub825 \ubc0f \ucd9c\ub825 \uc774\ub984 \uc815\uc758\noutput_path = \"mobilenetv2.onnx\"\nmodel_tensor = tf.convert_to_tensor(tf.ones([1, 224, 224, 3]), dtype=tf.float32)\n\n# ONNX \ubaa8\ub378\ub85c \ubcc0\ud658\nonnx_model = tf2onnx.convert.from_keras(model)\ntf2onnx.save_model(onnx_model, output_path)\n <\/code><\/pre>\n\uc774 \ucf54\ub4dc\ub294 Keras\ub97c \uc0ac\uc6a9\ud558\uc5ec MobileNetV2 \ubaa8\ub378\uc744 \uc0dd\uc131\ud558\uace0 \uc774\ub97c ONNX \ud30c\uc77c\ub85c \ubcc0\ud658\ud569\ub2c8\ub2e4. tf2onnx.convert.from_keras()<\/code> \uba54\uc18c\ub4dc\ub97c \uc0ac\uc6a9\ud558\uc5ec \ubaa8\ub378\uc744 ONNX\ub85c \ubcc0\ud658\ud558\uace0, tf2onnx.save_model()<\/code> \uba54\uc18c\ub4dc\ub85c \uc800\uc7a5\ud569\ub2c8\ub2e4.<\/p>\n<\/section>\n\n4. OpenCV\uc5d0\uc11c ONNX \ubaa8\ub378 \uc2e4\ud589\ud558\uae30<\/h2>\n
ONNX \ud30c\uc77c\uc774 \uc0dd\uc131\ub418\uba74, OpenCV\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc774 \ubaa8\ub378\uc744 \uc2e4\ud589\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. OpenCV\ub294 ONNX \ubaa8\ub378\uc744 \ub85c\ub4dc\ud558\uace0 \uc785\ub825 \uc774\ubbf8\uc9c0\ub97c \ubaa8\ub378\uc5d0 \uc804\ub2ec\ud558\uc5ec \uc608\uce21 \uacb0\uacfc\ub97c \uc5bb\ub294 \uae30\ub2a5\uc744 \uc81c\uacf5\ud569\ub2c8\ub2e4.<\/p>\n
4.1. OpenCV \uc124\uce58<\/h3>\n
OpenCV\ub97c \uc124\uce58\ud558\ub824\uba74 pip\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc124\uce58\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc544\ub798 \uba85\ub839\uc5b4\ub97c \ud1b5\ud574 OpenCV\ub97c \uc124\uce58\ud569\ub2c8\ub2e4.<\/p>\n
\npip install opencv-python\npip install opencv-python-headless\n <\/code><\/pre>\n4.2. ONNX \ubaa8\ub378\uc744 OpenCV\ub85c \ub85c\ub4dc\ud558\uace0 \uc2e4\ud589\ud558\uae30<\/h3>\n
\uc544\ub798 \uc608\uc81c\ub294 OpenCV\uc5d0\uc11c ONNX \ubaa8\ub378\uc744 \ub85c\ub4dc\ud558\uc5ec \uc774\ubbf8\uc9c0\ub97c \ubd84\ub958\ud558\ub294 \ubc29\ubc95\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4. \uc774 \uc608\uc81c\ub294 \uc774\uc804\uc5d0 \uc0dd\uc131\ud55c resnet18.onnx<\/code> \ud30c\uc77c\uc744 \uc0ac\uc6a9\ud569\ub2c8\ub2e4.<\/p>\n\nimport cv2\nimport numpy as np\n\n# ONNX \ubaa8\ub378 \ub85c\ub4dc\nnet = cv2.dnn.readNetFromONNX('resnet18.onnx')\n\n# \uc785\ub825 \uc774\ubbf8\uc9c0 \ub85c\ub4dc \ubc0f \uc804\ucc98\ub9ac\ninput_image = cv2.imread('input_image.jpg')\ninput_image = cv2.resize(input_image, (224, 224))\nblob = cv2.dnn.blobFromImage(input_image, 1.0, (224, 224), (104, 117, 123), swapRB=True, crop=False)\n\n# \ub124\ud2b8\uc6cc\ud06c\uc5d0 \uc785\ub825 \ub370\uc774\ud130 \uc124\uc815\nnet.setInput(blob)\n\n# \ubaa8\ub378 \uc2e4\ud589 \ubc0f \uacb0\uacfc \uc5bb\uae30\noutput = net.forward()\n\n# \uacb0\uacfc \ud6c4\ucc98\ub9ac (\uac00\uc7a5 \ub192\uc740 \ud655\ub960\uc758 \ud074\ub798\uc2a4 \ucc3e\uae30)\nclass_id = np.argmax(output[0])\nconfidence = output[0][class_id]\n\nprint(f\"Class ID: {class_id}, Confidence: {confidence:.2f}\")\n <\/code><\/pre>\n\uc704 \ucf54\ub4dc\uc5d0\uc11c\ub294 OpenCV\uc758 dnn<\/code> \ubaa8\ub4c8\uc744 \uc0ac\uc6a9\ud558\uc5ec ONNX \ubaa8\ub378\uc744 \ub85c\ub4dc\ud558\uace0, \uc785\ub825 \uc774\ubbf8\uc9c0\ub97c \uc804\ucc98\ub9ac\ud55c \ud6c4 \ubaa8\ub378\ub85c \uc608\uce21\uc744 \uc218\ud589\ud569\ub2c8\ub2e4. cv2.resize()<\/code> \uba54\uc18c\ub4dc\ub85c \uc774\ubbf8\uc9c0\ub97c 224×224 \ud06c\uae30\ub85c \uc870\uc815\ud558\uace0, cv2.dnn.blobFromImage()<\/code> \uba54\uc18c\ub4dc\ub85c \uc774\ubbf8\uc9c0\ub97c \ube14\ub86d \ud615\uc2dd\uc73c\ub85c \ubcc0\ud658\ud569\ub2c8\ub2e4. \ub9c8\uc9c0\ub9c9\uc73c\ub85c, net.forward()<\/code> \uba54\uc18c\ub4dc\ub97c \ud638\ucd9c\ud558\uc5ec \ubaa8\ub378\uc758 \uc608\uce21 \uacb0\uacfc\ub97c \uc5bb\uc2b5\ub2c8\ub2e4.<\/p>\n<\/section>\n\n5. \uacb0\ub860<\/h2>\n
ONNX \ud615\uc2dd\uc73c\ub85c\uc758 \ubaa8\ub378 \ubcc0\ud658\uacfc \uc774 \ubaa8\ub378\uc744 OpenCV\uc5d0\uc11c \uc2e4\ud589\ud558\ub294 \uacfc\uc815\uc740 \uac04\ub2e8\ud558\uc9c0\ub9cc \ub9e4\uc6b0 \uac15\ub825\ud55c \uae30\ub2a5\uc744 \uc81c\uacf5\ud569\ub2c8\ub2e4. \uc774\ub97c \ud1b5\ud574 \uc6b0\ub9ac\ub294 \ub2e4\uc591\ud55c \uba38\uc2e0\ub7ec\ub2dd \ud504\ub808\uc784\uc6cc\ud06c\uc5d0\uc11c \ud6c8\ub828\ub41c \ubaa8\ub378\uc744 \uc190\uc27d\uac8c \ud65c\uc6a9\ud560 \uc218 \uc788\uc73c\uba70, OpenCV\uc758 \ub2e4\uc591\ud55c \uc774\ubbf8\uc9c0 \ucc98\ub9ac \uae30\ub2a5\uc744 \ubaa8\ub378 \uacb0\uacfc\uc5d0 \uc801\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ubcf8 \uac15\uc88c\ub97c \ud1b5\ud574 \uc5ec\ub7ec\ubd84\uc774 OpenCV\uc640 ONNX\ub97c \ud568\uaed8 \uc0ac\uc6a9\ud558\ub294 \ubc29\ubc95\uc744 \uc774\ud574\ud558\uace0, \ubcf4\ub2e4 \ud6a8\uc728\uc801\uc778 \ub370\uc774\ud130 \ucc98\ub9ac\ub97c \ud560 \uc218 \uc788\uae30\ub97c \ubc14\ub78d\ub2c8\ub2e4.<\/p>\n<\/section>\n