{ "id": 39393, "title": "image pre-processing", "url": "/competitions/carvana-image-masking-challenge/discussion/39393", "author_name": "", "post_date": "2017-09-13T08:47:40.548828700Z", "votes": 6, "comment_count": 2, "views": 0, "content": "

here is opencv preprocessing. Note:

\n\n\n\n

links:

\n\n

http://docs.opencv.org/trunk/d7/d8f/tutorial_bioinspired_retina_illusion.html

\n\n

http://docs.opencv.org/trunk/d2/d94/bioinspired_retina.html

\n\n

https://pypkg.com/pypi/simretina/f/simretina/retina.py

\n\n
    img_file = CARVANA_DIR + '/images/%s/%s.jpg'%(folder,name)\n    image = cv2.imread(img_file) \n\n    retina = cv2.bioinspired.createRetina( inputSize = (width,height),)\n\n    retina.setupOPLandIPLParvoChannel(\n        1 , #para_dict[\"color_mode\"],\n        1 , #para_dict[\"normalise_output_parvo\"],\n        0.89 , #para_dict[\"photoreceptors_local_adaptation_sensitivity\"],\n        0.9 , #para_dict[\"photoreceptors_temporal_constant\"],\n        0.53 , #para_dict[\"photoreceptors_spatial_constant\"],\n        0.3 , #para_dict[\"horizontal_cells_gain\"],\n        0.5 , #para_dict[\"hcells_temporal_constant\"],\n        7 , #para_dict[\"hcells_spatial_constant\"],\n        0.89 , #para_dict[\"ganglion_cells_sensitivity\"]\n    )\n\n    retina.setupIPLMagnoChannel(\n        1 , #para_dict[\"normalise_output_magno\"],\n        0 , #para_dict[\"parasol_cells_beta\"],\n        0 , #para_dict[\"parasol_cells_tau\"],\n        7 , #para_dict[\"parasol_cells_k\"],\n        2 , #para_dict[\"amacrin_cells_temporal_cut_frequency\"],\n        0.95 , #para_dict[\"v0_compression_parameter\"],\n        0 ,    #para_dict[\"local_adapt_integration_tau\"],\n        7 ,    #para_dict[\"local_adapt_integration_k\"]\n    )\n\n    for i in range(20):\n       print (i)\n       retina.run(image)\n\n\n    retinaOut_parvo = retina.getParvo()\n    im_show('image', image,  resize=1)\n    im_show('retinaOut_parvo',  retinaOut_parvo,  resize=1)\n    cv2.waitKey(0)\n
", "messages": [ { "id": "220840", "postDate": "09/13/2017 08:47:40", "content": "

here is opencv preprocessing. Note:

\n\n\n\n

links:

\n\n

http://docs.opencv.org/trunk/d7/d8f/tutorial_bioinspired_retina_illusion.html

\n\n

http://docs.opencv.org/trunk/d2/d94/bioinspired_retina.html

\n\n

https://pypkg.com/pypi/simretina/f/simretina/retina.py

\n\n
    img_file = CARVANA_DIR + '/images/%s/%s.jpg'%(folder,name)\n    image = cv2.imread(img_file) \n\n    retina = cv2.bioinspired.createRetina( inputSize = (width,height),)\n\n    retina.setupOPLandIPLParvoChannel(\n        1 , #para_dict[\"color_mode\"],\n        1 , #para_dict[\"normalise_output_parvo\"],\n        0.89 , #para_dict[\"photoreceptors_local_adaptation_sensitivity\"],\n        0.9 , #para_dict[\"photoreceptors_temporal_constant\"],\n        0.53 , #para_dict[\"photoreceptors_spatial_constant\"],\n        0.3 , #para_dict[\"horizontal_cells_gain\"],\n        0.5 , #para_dict[\"hcells_temporal_constant\"],\n        7 , #para_dict[\"hcells_spatial_constant\"],\n        0.89 , #para_dict[\"ganglion_cells_sensitivity\"]\n    )\n\n    retina.setupIPLMagnoChannel(\n        1 , #para_dict[\"normalise_output_magno\"],\n        0 , #para_dict[\"parasol_cells_beta\"],\n        0 , #para_dict[\"parasol_cells_tau\"],\n        7 , #para_dict[\"parasol_cells_k\"],\n        2 , #para_dict[\"amacrin_cells_temporal_cut_frequency\"],\n        0.95 , #para_dict[\"v0_compression_parameter\"],\n        0 ,    #para_dict[\"local_adapt_integration_tau\"],\n        7 ,    #para_dict[\"local_adapt_integration_k\"]\n    )\n\n    for i in range(20):\n       print (i)\n       retina.run(image)\n\n\n    retinaOut_parvo = retina.getParvo()\n    im_show('image', image,  resize=1)\n    im_show('retinaOut_parvo',  retinaOut_parvo,  resize=1)\n    cv2.waitKey(0)\n
", "rawMarkdown": "here is opencv preprocessing. Note:\n\n - it take a long time to make these results\n\n - I am not sure if they are useful or not\n\nlinks:\n\nhttp://docs.opencv.org/trunk/d7/d8f/tutorial_bioinspired_retina_illusion.html\n\nhttp://docs.opencv.org/trunk/d2/d94/bioinspired_retina.html\n\nhttps://pypkg.com/pypi/simretina/f/simretina/retina.py\n\n img_file = CARVANA_DIR + '/images/%s/%s.jpg'%(folder,name)\n image = cv2.imread(img_file) \n\n retina = cv2.bioinspired.createRetina( inputSize = (width,height),)\n\n retina.setupOPLandIPLParvoChannel(\n 1 , #para_dict[\"color_mode\"],\n 1 , #para_dict[\"normalise_output_parvo\"],\n 0.89 , #para_dict[\"photoreceptors_local_adaptation_sensitivity\"],\n 0.9 , #para_dict[\"photoreceptors_temporal_constant\"],\n 0.53 , #para_dict[\"photoreceptors_spatial_constant\"],\n 0.3 , #para_dict[\"horizontal_cells_gain\"],\n 0.5 , #para_dict[\"hcells_temporal_constant\"],\n 7 , #para_dict[\"hcells_spatial_constant\"],\n 0.89 , #para_dict[\"ganglion_cells_sensitivity\"]\n )\n\n retina.setupIPLMagnoChannel(\n 1 , #para_dict[\"normalise_output_magno\"],\n 0 , #para_dict[\"parasol_cells_beta\"],\n 0 , #para_dict[\"parasol_cells_tau\"],\n 7 , #para_dict[\"parasol_cells_k\"],\n 2 , #para_dict[\"amacrin_cells_temporal_cut_frequency\"],\n 0.95 , #para_dict[\"v0_compression_parameter\"],\n 0 , #para_dict[\"local_adapt_integration_tau\"],\n 7 , #para_dict[\"local_adapt_integration_k\"]\n )\n\n for i in range(20):\n print (i)\n retina.run(image)\n\n\n retinaOut_parvo = retina.getParvo()\n im_show('image', image, resize=1)\n im_show('retinaOut_parvo', retinaOut_parvo, resize=1)\n cv2.waitKey(0)", "votes": null }, { "id": "220842", "postDate": "09/13/2017 08:49:46", "content": "

this compares the gradient orientation images of the hq and old images.

\n\n
def draw_hsv(flow):\n   h, w = flow.shape[:2]\n   fx, fy = flow[:,:,0], flow[:,:,1]\n   ang = np.arctan2(fy, fx) + np.pi\n   v = np.sqrt(fx*fx+fy*fy)\n   hsv = np.zeros((h, w, 3), np.uint8)\n   hsv[...,0] = ang*(180/np.pi/2)\n   hsv[...,1] = 255\n   hsv[...,2] = np.minimum(v*4, 255)\n   bgr = cv2.cvtColor(hsv, cv2.COLOR_HSV2BGR)\nreturn bgr\n\ndef run_check_flow():\n\n   folder ='test_hq1024x1024'\n   #folder ='test512x512'\n   #folder ='test'\n   #folder ='test_hq'\n   #name = 'a67333bd4065_09'  #'0c56c040a690_09'\n   name = '1b218148cb02_09'  #'0c56c040a690_09'\n\n   img_file = CARVANA_DIR + '/images/%s/%s.jpg'%(folder,name)\n   image = cv2.imread(img_file)\n\n   gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)\n   height,width,C = image.shape\n   dx =np.zeros((height,width,1),np.float32)\n   dy =np.zeros((height,width,1),np.float32)\n\n   dx[:,1:,0] = gray[:,1:]-gray[:,:-1]\n   dy[1:,:,0] = gray[1:]-gray[:-1]\n\n   flow = np.dstack((dx,dy))\n\n   f = draw_hsv(flow)\n   cv2.imwrite('/root/share/project/kaggle-carvana-cars/results/processing/test_hq.png',f)\n   im_show('flow',  f,  resize=1)\n   im_show('dx',  np.abs(dx),  resize=1)\n   im_show('dy',  np.abs(dy),  resize=1)\n   cv2.waitKey(0)\n
", "rawMarkdown": "this compares the gradient orientation images of the hq and old images.\n\n\n def draw_hsv(flow):\n h, w = flow.shape[:2]\n fx, fy = flow[:,:,0], flow[:,:,1]\n ang = np.arctan2(fy, fx) + np.pi\n v = np.sqrt(fx*fx+fy*fy)\n hsv = np.zeros((h, w, 3), np.uint8)\n hsv[...,0] = ang*(180/np.pi/2)\n hsv[...,1] = 255\n hsv[...,2] = np.minimum(v*4, 255)\n bgr = cv2.cvtColor(hsv, cv2.COLOR_HSV2BGR)\n return bgr\n\n def run_check_flow():\n\n folder ='test_hq1024x1024'\n #folder ='test512x512'\n #folder ='test'\n #folder ='test_hq'\n #name = 'a67333bd4065_09' #'0c56c040a690_09'\n name = '1b218148cb02_09' #'0c56c040a690_09'\n\n img_file = CARVANA_DIR + '/images/%s/%s.jpg'%(folder,name)\n image = cv2.imread(img_file)\n\n gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)\n height,width,C = image.shape\n dx =np.zeros((height,width,1),np.float32)\n dy =np.zeros((height,width,1),np.float32)\n\n dx[:,1:,0] = gray[:,1:]-gray[:,:-1]\n dy[1:,:,0] = gray[1:]-gray[:-1]\n\n flow = np.dstack((dx,dy))\n\n f = draw_hsv(flow)\n cv2.imwrite('/root/share/project/kaggle-carvana-cars/results/processing/test_hq.png',f)\n im_show('flow', f, resize=1)\n im_show('dx', np.abs(dx), resize=1)\n im_show('dy', np.abs(dy), resize=1)\n cv2.waitKey(0)", "votes": null }, { "id": "220873", "postDate": "09/13/2017 10:59:18", "content": "

I did a similar thing using histogram equalization. It helped with convergence speed but did not increase overall accuracy in a noticeable way.

\n\n
import cv2\nimport os\nfrom scipy import ndimage\nimport matplotlib\nmatplotlib.use('TkAgg')\nfrom matplotlib import pyplot as plt\nimport numpy as np\n\ndef equalize(image_path, output_folder, testing=False):\n    if not testing:\n        new_clean_name = os.path.join(output_folder, os.path.basename(image_path))\n\n        if os.path.exists(new_clean_name):\n            return\n\n    #-----Reading the image-----------------------------------------------------\n    img = cv2.imread(image_path, 1)\n\n    #-----Converting image to LAB Color model----------------------------------- \n    lab= cv2.cvtColor(img, cv2.COLOR_BGR2LAB)\n    # cv2.imshow(\"lab\",lab)\n\n    #-----Splitting the LAB image to different channels-------------------------\n    l, a, b = cv2.split(lab)\n\n    #-----Applying CLAHE to L-channel-------------------------------------------\n    clahe = cv2.createCLAHE(clipLimit=6.0, tileGridSize=(64,64))\n    cl = clahe.apply(l)\n\n    #-----Merge the CLAHE enhanced L-channel with the a and b channel-----------\n    limg = cv2.merge((cl,a,b))\n\n    #-----Converting image from LAB Color model to RGB model--------------------\n    final = cv2.cvtColor(limg, cv2.COLOR_LAB2BGR)\n\n    # cv2.imshow('final', final)\n    # cv2.imwrite(new_edge_name, final_edges)\n    if not testing:\n        cv2.imwrite(new_clean_name, final)\n    else:\n        plt.imshow(cv2.cvtColor(limg, cv2.COLOR_LAB2RGB))\n        plt.show()\n
", "rawMarkdown": "I did a similar thing using histogram equalization. It helped with convergence speed but did not increase overall accuracy in a noticeable way.\n\n import cv2\n import os\n from scipy import ndimage\n import matplotlib\n matplotlib.use('TkAgg')\n from matplotlib import pyplot as plt\n import numpy as np\n \n def equalize(image_path, output_folder, testing=False):\n if not testing:\n new_clean_name = os.path.join(output_folder, os.path.basename(image_path))\n \n if os.path.exists(new_clean_name):\n return\n \n #-----Reading the image-----------------------------------------------------\n img = cv2.imread(image_path, 1)\n \n #-----Converting image to LAB Color model----------------------------------- \n lab= cv2.cvtColor(img, cv2.COLOR_BGR2LAB)\n # cv2.imshow(\"lab\",lab)\n \n #-----Splitting the LAB image to different channels-------------------------\n l, a, b = cv2.split(lab)\n \n #-----Applying CLAHE to L-channel-------------------------------------------\n clahe = cv2.createCLAHE(clipLimit=6.0, tileGridSize=(64,64))\n cl = clahe.apply(l)\n \n #-----Merge the CLAHE enhanced L-channel with the a and b channel-----------\n limg = cv2.merge((cl,a,b))\n \n #-----Converting image from LAB Color model to RGB model--------------------\n final = cv2.cvtColor(limg, cv2.COLOR_LAB2BGR)\n \n # cv2.imshow('final', final)\n # cv2.imwrite(new_edge_name, final_edges)\n if not testing:\n cv2.imwrite(new_clean_name, final)\n else:\n plt.imshow(cv2.cvtColor(limg, cv2.COLOR_LAB2RGB))\n plt.show()", "votes": null } ], "comments": [ { "id": 220842, "author_name": "hengck23", "author_url": "", "post_date": "09/13/2017 08:49:46", "content": "

this compares the gradient orientation images of the hq and old images.

\n\n
def draw_hsv(flow):\n   h, w = flow.shape[:2]\n   fx, fy = flow[:,:,0], flow[:,:,1]\n   ang = np.arctan2(fy, fx) + np.pi\n   v = np.sqrt(fx*fx+fy*fy)\n   hsv = np.zeros((h, w, 3), np.uint8)\n   hsv[...,0] = ang*(180/np.pi/2)\n   hsv[...,1] = 255\n   hsv[...,2] = np.minimum(v*4, 255)\n   bgr = cv2.cvtColor(hsv, cv2.COLOR_HSV2BGR)\nreturn bgr\n\ndef run_check_flow():\n\n   folder ='test_hq1024x1024'\n   #folder ='test512x512'\n   #folder ='test'\n   #folder ='test_hq'\n   #name = 'a67333bd4065_09'  #'0c56c040a690_09'\n   name = '1b218148cb02_09'  #'0c56c040a690_09'\n\n   img_file = CARVANA_DIR + '/images/%s/%s.jpg'%(folder,name)\n   image = cv2.imread(img_file)\n\n   gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)\n   height,width,C = image.shape\n   dx =np.zeros((height,width,1),np.float32)\n   dy =np.zeros((height,width,1),np.float32)\n\n   dx[:,1:,0] = gray[:,1:]-gray[:,:-1]\n   dy[1:,:,0] = gray[1:]-gray[:-1]\n\n   flow = np.dstack((dx,dy))\n\n   f = draw_hsv(flow)\n   cv2.imwrite('/root/share/project/kaggle-carvana-cars/results/processing/test_hq.png',f)\n   im_show('flow',  f,  resize=1)\n   im_show('dx',  np.abs(dx),  resize=1)\n   im_show('dy',  np.abs(dy),  resize=1)\n   cv2.waitKey(0)\n
", "votes": null, "replies": [] }, { "id": 220873, "author_name": "adamhart", "author_url": "", "post_date": "09/13/2017 10:59:18", "content": "

I did a similar thing using histogram equalization. It helped with convergence speed but did not increase overall accuracy in a noticeable way.

\n\n
import cv2\nimport os\nfrom scipy import ndimage\nimport matplotlib\nmatplotlib.use('TkAgg')\nfrom matplotlib import pyplot as plt\nimport numpy as np\n\ndef equalize(image_path, output_folder, testing=False):\n    if not testing:\n        new_clean_name = os.path.join(output_folder, os.path.basename(image_path))\n\n        if os.path.exists(new_clean_name):\n            return\n\n    #-----Reading the image-----------------------------------------------------\n    img = cv2.imread(image_path, 1)\n\n    #-----Converting image to LAB Color model----------------------------------- \n    lab= cv2.cvtColor(img, cv2.COLOR_BGR2LAB)\n    # cv2.imshow(\"lab\",lab)\n\n    #-----Splitting the LAB image to different channels-------------------------\n    l, a, b = cv2.split(lab)\n\n    #-----Applying CLAHE to L-channel-------------------------------------------\n    clahe = cv2.createCLAHE(clipLimit=6.0, tileGridSize=(64,64))\n    cl = clahe.apply(l)\n\n    #-----Merge the CLAHE enhanced L-channel with the a and b channel-----------\n    limg = cv2.merge((cl,a,b))\n\n    #-----Converting image from LAB Color model to RGB model--------------------\n    final = cv2.cvtColor(limg, cv2.COLOR_LAB2BGR)\n\n    # cv2.imshow('final', final)\n    # cv2.imwrite(new_edge_name, final_edges)\n    if not testing:\n        cv2.imwrite(new_clean_name, final)\n    else:\n        plt.imshow(cv2.cvtColor(limg, cv2.COLOR_LAB2RGB))\n        plt.show()\n
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "220840": "here is opencv preprocessing. Note:\n\n - it take a long time to make these results\n\n - I am not sure if they are useful or not\n\nlinks:\n\nhttp://docs.opencv.org/trunk/d7/d8f/tutorial_bioinspired_retina_illusion.html\n\nhttp://docs.opencv.org/trunk/d2/d94/bioinspired_retina.html\n\nhttps://pypkg.com/pypi/simretina/f/simretina/retina.py\n\n img_file = CARVANA_DIR + '/images/%s/%s.jpg'%(folder,name)\n image = cv2.imread(img_file) \n\n retina = cv2.bioinspired.createRetina( inputSize = (width,height),)\n\n retina.setupOPLandIPLParvoChannel(\n 1 , #para_dict[\"color_mode\"],\n 1 , #para_dict[\"normalise_output_parvo\"],\n 0.89 , #para_dict[\"photoreceptors_local_adaptation_sensitivity\"],\n 0.9 , #para_dict[\"photoreceptors_temporal_constant\"],\n 0.53 , #para_dict[\"photoreceptors_spatial_constant\"],\n 0.3 , #para_dict[\"horizontal_cells_gain\"],\n 0.5 , #para_dict[\"hcells_temporal_constant\"],\n 7 , #para_dict[\"hcells_spatial_constant\"],\n 0.89 , #para_dict[\"ganglion_cells_sensitivity\"]\n )\n\n retina.setupIPLMagnoChannel(\n 1 , #para_dict[\"normalise_output_magno\"],\n 0 , #para_dict[\"parasol_cells_beta\"],\n 0 , #para_dict[\"parasol_cells_tau\"],\n 7 , #para_dict[\"parasol_cells_k\"],\n 2 , #para_dict[\"amacrin_cells_temporal_cut_frequency\"],\n 0.95 , #para_dict[\"v0_compression_parameter\"],\n 0 , #para_dict[\"local_adapt_integration_tau\"],\n 7 , #para_dict[\"local_adapt_integration_k\"]\n )\n\n for i in range(20):\n print (i)\n retina.run(image)\n\n\n retinaOut_parvo = retina.getParvo()\n im_show('image', image, resize=1)\n im_show('retinaOut_parvo', retinaOut_parvo, resize=1)\n cv2.waitKey(0)", "220842": "this compares the gradient orientation images of the hq and old images.\n\n\n def draw_hsv(flow):\n h, w = flow.shape[:2]\n fx, fy = flow[:,:,0], flow[:,:,1]\n ang = np.arctan2(fy, fx) + np.pi\n v = np.sqrt(fx*fx+fy*fy)\n hsv = np.zeros((h, w, 3), np.uint8)\n hsv[...,0] = ang*(180/np.pi/2)\n hsv[...,1] = 255\n hsv[...,2] = np.minimum(v*4, 255)\n bgr = cv2.cvtColor(hsv, cv2.COLOR_HSV2BGR)\n return bgr\n\n def run_check_flow():\n\n folder ='test_hq1024x1024'\n #folder ='test512x512'\n #folder ='test'\n #folder ='test_hq'\n #name = 'a67333bd4065_09' #'0c56c040a690_09'\n name = '1b218148cb02_09' #'0c56c040a690_09'\n\n img_file = CARVANA_DIR + '/images/%s/%s.jpg'%(folder,name)\n image = cv2.imread(img_file)\n\n gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)\n height,width,C = image.shape\n dx =np.zeros((height,width,1),np.float32)\n dy =np.zeros((height,width,1),np.float32)\n\n dx[:,1:,0] = gray[:,1:]-gray[:,:-1]\n dy[1:,:,0] = gray[1:]-gray[:-1]\n\n flow = np.dstack((dx,dy))\n\n f = draw_hsv(flow)\n cv2.imwrite('/root/share/project/kaggle-carvana-cars/results/processing/test_hq.png',f)\n im_show('flow', f, resize=1)\n im_show('dx', np.abs(dx), resize=1)\n im_show('dy', np.abs(dy), resize=1)\n cv2.waitKey(0)", "220873": "I did a similar thing using histogram equalization. It helped with convergence speed but did not increase overall accuracy in a noticeable way.\n\n import cv2\n import os\n from scipy import ndimage\n import matplotlib\n matplotlib.use('TkAgg')\n from matplotlib import pyplot as plt\n import numpy as np\n \n def equalize(image_path, output_folder, testing=False):\n if not testing:\n new_clean_name = os.path.join(output_folder, os.path.basename(image_path))\n \n if os.path.exists(new_clean_name):\n return\n \n #-----Reading the image-----------------------------------------------------\n img = cv2.imread(image_path, 1)\n \n #-----Converting image to LAB Color model----------------------------------- \n lab= cv2.cvtColor(img, cv2.COLOR_BGR2LAB)\n # cv2.imshow(\"lab\",lab)\n \n #-----Splitting the LAB image to different channels-------------------------\n l, a, b = cv2.split(lab)\n \n #-----Applying CLAHE to L-channel-------------------------------------------\n clahe = cv2.createCLAHE(clipLimit=6.0, tileGridSize=(64,64))\n cl = clahe.apply(l)\n \n #-----Merge the CLAHE enhanced L-channel with the a and b channel-----------\n limg = cv2.merge((cl,a,b))\n \n #-----Converting image from LAB Color model to RGB model--------------------\n final = cv2.cvtColor(limg, cv2.COLOR_LAB2BGR)\n \n # cv2.imshow('final', final)\n # cv2.imwrite(new_edge_name, final_edges)\n if not testing:\n cv2.imwrite(new_clean_name, final)\n else:\n plt.imshow(cv2.cvtColor(limg, cv2.COLOR_LAB2RGB))\n plt.show()" }, "source": "meta" }