Upgraded to python3, and added setup.py and installation instructions

README.md

@@ -1,25 +1,56 @@
 train-detector
 ==============
 
-This repository contains scripts that will help train a license plate detector for a particular region.  Your trained region detector can then be used in OpenALPR.
-
-The license plate region detector uses the Local Binary Pattern (LBP) algorithm.  In order to train the detector, you will need many positive and negative images.  This repository already contains a collection of negative images.  You will need to add your own positive images.
-
-To get started, you will first need many cropped plate images containing positive license plate matches.  Please see the "eu" positive image folder in this repository to understand the types of plate images required. 
-
-The [Plate Tagger Utility](https://github.com/openalpr/plate_tagger)  is helpful to tag the plate locations.  After tagging the plates you can run the "crop_plates.py" function to extract the crops from the input images at your target aspect ratio.
-
-After you've collected many (hundreds to thousands) of positive plate images, the next step is to train the detector.  First you must configure the training script to use the correct dimensions.
-
-Edit the prep.py script and change the WIDTH, HEIGHT, and COUNTRY variables to match the country that you are training.  The width and height should be proportional to the plate size (slightly larger is OK).  A total pixel area of around 650 seems to work best.  Also, adjust the path to your OpenCV libraries, if that needs to be changed.
+This repository contains scripts that will help train a license plate detector
+for a particular region.  Your trained region detector can then be used in
+OpenALPR.
+
+The license plate region detector uses the Local Binary Pattern (LBP)
+algorithm.  In order to train the detector, you will need many positive and
+negative images.  This repository already contains a collection of negative
+images.  You will need to add your own positive images.
+
+To get started, you will first need many cropped plate images containing
+positive license plate matches.  Please see the "eu" positive image folder in
+this repository to understand the types of plate images required. 
+
+The [Plate Tagger Utility](https://github.com/openalpr/plate_tagger)  is
+helpful to tag the plate locations.  After tagging the plates you can run the
+"crop_plates.py" function to extract the crops from the input images at your
+target aspect ratio.
+
+```
+python3 -m venv td
+source td/bin/activate
+python -m pip install --editable .
+python3 crop_plates.py --input_dir /tmp/pool --out_dir ~/work/ANPR-RevenueNSW/data/cropped_plates 
+```
+
+After you've collected many (hundreds to thousands) of positive plate images,
+the next step is to train the detector.  First you must configure the training
+script to use the correct dimensions.
+
+Edit the prep.py script and change the WIDTH, HEIGHT, and COUNTRY variables to
+match the country that you are training.  The width and height should be
+proportional to the plate size (slightly larger is OK).  A total pixel area of
+around 650 seems to work best.  Also, adjust the path to your OpenCV libraries,
+if that needs to be changed.
 
 Once you are ready to start training, enter the following commands:
 
-  - rm ./out/*    (clear the out folder in case it has data from previous runs)
-  - ./prep.py neg
-  - ./prep.py pos
-  - ./prep.py train
-  - Copy the output from the above command onto the command line.  You should adjust the numStages to a smaller value (usually 12 stages works well, but it will depend on your input images).  You may also need to adjust the numPos value to a smaller number in order to complete the training.
+```
+rm ./out/*    (clear the out folder in case it has data from previous runs)
+python3 prep.py neg
+python3 prep.py pos
+python3 prep.py train
+```
+
+Copy the output from the above command onto the command line.  You should
+adjust the numStages to a smaller value (usually 12 stages works well, but it
+will depend on your input images).  You may also need to adjust the numPos
+value to a smaller number in order to complete the training.
 
 
-Copy the out/cascade.xml file to your OpenALPR runtime directory (runtime_data/region/[countrycode].xml).  You should now be able to use the region for plate detection.
+Copy the out/cascade.xml file to your OpenALPR runtime directory
+(runtime_data/region/[countrycode].xml).  You should now be able to use the
+region for plate detection.

crop_plates.py

@@ -4,7 +4,7 @@
 import sys
 import json
 import math
-import cv, cv2
+import cv2
 import numpy as np
 import copy
 import yaml
@@ -22,16 +22,12 @@
 parser.add_argument( "--zoom_out_percent", dest="zoom_out_percent", action="store", type=float, default=1.25, 
                   help="Percent multiplier to zoom out before cropping" )
 
-parser.add_argument( "--plate_width", dest="plate_width", action="store", type=float, required=True, 
-                  help="Desired aspect ratio width" )
-parser.add_argument( "--plate_height", dest="plate_height", action="store", type=float, required=True, 
-                  help="Desired aspect ratio height" )
 
 options = parser.parse_args()
 
 
 if not os.path.isdir(options.input_dir):
-    print "input_dir (%s) doesn't exist"
+    print("input_dir (%s) doesn't exist")
     sys.exit(1)
 
 
@@ -40,82 +36,6 @@
 
 
 
-def get_box(x1, y1, x2, y2, x3, y3, x4, y4):
-    height1 = int(round(math.sqrt((x1-x4)*(x1-x4) + (y1-y4)*(y1-y4))))
-    height2 = int(round(math.sqrt((x3-x2)*(x3-x2) + (y3-y2)*(y3-y2))))
-
-    height = height1
-    if height2 > height:
-        height = height2
-
-    # add 25% to the height
-    height *= options.zoom_out_percent
-    #height += (height * .05)
-
-    #print "Height: %d - %d" % (height1, height2)
-
-
-    points = [(x1,y1), (x2,y2), (x3,y3), (x4,y4)]
-    moment = cv.Moments(points)
-    centerx = int(round(moment.m10/moment.m00))
-    centery = int(round(moment.m01/moment.m00))
-
-
-    training_aspect = options.plate_width / options.plate_height
-    width = int(round(training_aspect * height))
-
-    # top_left = ( int(centerx - (width / 2)), int(centery - (height / 2)))
-    # bottom_right = ( int(centerx + (width / 2)), int(centery + (height / 2)))
-
-    top_left_x = int(round(centerx - (width / 2)))
-    top_left_y = int(round(centery - (height / 2)))
-
-    return (top_left_x, top_left_y, width, int(round(height)))
-
-def crop_rect(big_image, x,y,width,height):
-    # Crops the rectangle from the big image and returns a cropped image
-    # Special care is taken to avoid cropping beyond the edge of the image.
-    # It fills this area in with random pixels
-
-    (big_height, big_width, channels) = big_image.shape
-    if x >= 0 and y >= 0 and (y+height) < big_height and (x+width) < big_width:
-        crop_img = img[y:y+height, x:x+width]
-    else:
-        #print "Performing partial crop"
-        #print "x: %d  y: %d  width: %d  height: %d" % (x,y,width,height)
-        #print "big_width: %d  big_height: %d" % (big_width, big_height)
-        crop_img = np.zeros((height, width, 3), np.uint8)
-        cv2.randu(crop_img, (0,0,0), (255,255,255))
-
-        offset_x = 0
-        offset_y = 0
-        if x < 0:
-            offset_x = -1 * x
-            x = 0
-            width -= offset_x
-        if y < 0:
-            offset_y = -1 * y
-            y = 0
-            height -= offset_y
-        if (x+width) >= big_width:
-            offset_x = 0
-            width = big_width - x
-        if (y+height) >= big_height:
-            offset_y = 0
-            height = big_height - y
-
-        #print "offset_x: %d  offset_y: %d, width: %d, height: %d" % (offset_x, offset_y, width, height)
-
-        original_crop =  img[y:y+height-1, x:x+width-1]
-        (small_image_height, small_image_width, channels) = original_crop.shape
-        #print "Small shape: %dx%d" % (small_image_width, small_image_height)
-        # Draw the small image onto the large image
-        crop_img[offset_y:offset_y+small_image_height, offset_x:offset_x+small_image_width] = original_crop
-
-
-    #cv2.imshow("Test", crop_img)
-    return crop_img
-
 count = 1
 yaml_files = []
 for in_file in os.listdir(options.input_dir):
@@ -128,7 +48,7 @@ def crop_rect(big_image, x,y,width,height):
 for yaml_file in yaml_files:
 
 
-    print "Processing: " + yaml_file + " (" + str(count) + "/" + str(len(yaml_files)) + ")"
+    print("Processing: " + yaml_file + " (" + str(count) + "/" + str(len(yaml_files)) + ")")
     count += 1
 
 
@@ -142,7 +62,7 @@ def crop_rect(big_image, x,y,width,height):
     # Skip missing images
     full_image_path = os.path.join(options.input_dir, image)
     if not os.path.isfile(full_image_path):
-        print "Could not find image file %s, skipping" % (full_image_path)
+        print("Could not find image file %s, skipping" % (full_image_path))
         continue
 
 
@@ -151,16 +71,17 @@ def crop_rect(big_image, x,y,width,height):
     for i in range(0, len(cc)):
         cc[i] = int(cc[i])
 
-    box = get_box(cc[0], cc[1], cc[2], cc[3], cc[4], cc[5], cc[6], cc[7])
-
-
     img = cv2.imread(full_image_path)
-    crop = crop_rect(img, box[0], box[1], box[2], box[3])
+    mask = np.zeros(img.shape[0:2], dtype=np.uint8)
+    points = np.array([[[cc[0],cc[1]],[cc[2], cc[3]], [cc[4],cc[5]], [cc[6],cc[7]]]])
+
+    cv2.drawContours(mask, [points], -1, (255,255,255), -1, cv2.LINE_AA)
 
-    # cv2.imshow("test", crop)
-    # cv2.waitKey(0)
+    res = cv2.bitwise_and(img,img,mask = mask)
+    rect = cv2.boundingRect(points) # returns (x,y,w,h) of the rect
+    cropped = res[rect[1]: rect[1] + rect[3], rect[0]: rect[0] + rect[2]]
+    out_crop_path = os.path.join(options.out_dir, os.path.basename(yaml_without_ext) + ".jpg")
+    cv2.imwrite(out_crop_path, cropped )
 
-    out_crop_path = os.path.join(options.out_dir, yaml_without_ext + ".jpg")
-    cv2.imwrite(out_crop_path, crop )
 
-print "%d Cropped images are located in %s" % (count-1, options.out_dir)
+print("%d Cropped images are located in %s" % (count-1, options.out_dir))

prep.py

@@ -10,7 +10,7 @@
 
 WIDTH=36
 HEIGHT=18
-COUNTRY='us'
+COUNTRY='nsw'
 
 #WIDTH=52
 #HEIGHT=13
@@ -21,10 +21,12 @@
 #COUNTRY='br'
 
 #constants
-OPENCV_DIR= '/home/mhill/projects/alpr/libraries/opencv/bin'
-SAMPLE_CREATOR = OPENCV_DIR + '/opencv_createsamples'
+OPENCV_DIR = "/usr/local/Cellar/opencv/4.2.0_3/bin/"
+SAMPLE_CREATOR = "/usr/local/Cellar/opencv@2/2.4.13.7_7/bin/opencv_createsamples"
+TRAINCASCADE = "/usr/local/Cellar/opencv@2/2.4.13.7_7/bin/opencv_traincascade"
 
-BASE_DIR            = '/home/mhill/projects/alpr/samples/training/'
+
+BASE_DIR            = './'
 
 OUTPUT_DIR          = BASE_DIR + "out/"
 INPUT_NEGATIVE_DIR  = BASE_DIR + 'raw-neg/'
@@ -44,12 +46,15 @@
 
 
 def print_usage():
-    print "Usage: prep.py [Operation]"
-    print "   -- Operations --"
-    print "  neg        -- Prepares the negative samples list"
-    print "  pos        -- Copies all the raw positive files to a opencv vector"
-    print "  showpos    -- Shows the positive samples that were created"
-    print "  train      -- Outputs the command for the Cascade Training algorithm"
+    usage = '''
+        Usage: prep.py [Operation]
+           -- Operations --
+          neg        -- Prepares the negative samples list
+          pos        -- Copies all the raw positive files to a opencv vector
+          showpos    -- Shows the positive samples that were created
+          train      -- Outputs the command for the Cascade Training algorithm
+        '''
+    print(usage)
 
 def file_len(fname):
     with open(fname) as f:
@@ -71,7 +76,7 @@ def file_len(fname):
 
 
 if command == "neg":
-    print "Neg"
+    print("Neg")
 
     # Get rid of any spaces
     for neg_file in os.listdir(INPUT_NEGATIVE_DIR):
@@ -97,7 +102,7 @@ def file_len(fname):
     f.close()
 
 elif command == "pos":
-    print "Pos"
+    print("Pos")
     info_arg = '-info %s' % (POSITIVE_INFO_FILE)
 
     # Copy all files in the raw directory and build an info file
@@ -133,52 +138,50 @@ def file_len(fname):
 
         if filename.endswith(".txt"):
             continue
-	try:
-		img = Image.open(OUTPUT_POSITIVE_DIR + filename)
-
-		# get the image's width and height in pixels
-		width, height = img.size
-		f.write(filename + " 1 0 0 " + str(width) + " " + str(height) + '\n')
+        try:
+            img = Image.open(OUTPUT_POSITIVE_DIR + filename)
+            # get the image's width and height in pixels
+            width, height = img.size
+            f.write(filename + " 1 0 0 " + str(width) + " " + str(height) + '\n')
 
-		total_pics = total_pics + 1
-	except IOError:
-		print "Exception reading image file: " + filename
+            total_pics = total_pics + 1
+        except IOError:
+            print("Exception reading image file: " + filename)
 
     f.close()
 
 
 
 
     # Collapse the samples into a vector file
-    execStr = '%s/opencv_createsamples %s %s %s -num %d' % (OPENCV_DIR, vector_arg, width_height_arg, info_arg, total_pics )
-    print execStr
+    execStr = '%s %s %s %s -num %d' % (SAMPLE_CREATOR, vector_arg, width_height_arg, info_arg, total_pics )
+    print(execStr)
 
     os.system(execStr)
     #opencv_createsamples -info ./positive.txt -vec ../positive/vecfile.vec -w 120 -h 60 -bg ../negative/PentagonCityParkingGarage21.jpg -num 100
 
 
 elif command == "showpos":
-    print "SHOW"
-    execStr = '%s/opencv_createsamples -vec %s -w %d -h %d' % (OPENCV_DIR, VEC_FILE, WIDTH, HEIGHT )
-    print execStr
+    print("SHOW")
+    execStr = '%s -vec %s -w %d -h %d' % (SAMPLE_CREATOR, VEC_FILE, WIDTH, HEIGHT )
+    print(execStr)
     os.system(execStr)
     #opencv_createsamples -vec ../positive/vecfile.vec -w 120 -h 60
 elif command == "train":
-    print "TRAIN"
+    print("TRAIN")
 
     data_arg = '-data %s/' % (OUTPUT_DIR)
     bg_arg = '-bg %s' % (NEGATIVE_INFO_FILE)
 
     try:
-	num_pos_samples = file_len(POSITIVE_INFO_FILE)
+        num_pos_samples = file_len(POSITIVE_INFO_FILE)
     except:
-	num_pos_samples = -1
+        num_pos_samples = -1
     num_neg_samples = file_len(NEGATIVE_INFO_FILE)
 
-    execStr = '%s/opencv_traincascade %s %s %s %s -numPos %d -numNeg %d -maxFalseAlarmRate 0.45 -featureType LBP -numStages 13' % (OPENCV_DIR, data_arg, vector_arg, bg_arg, width_height_arg, num_pos_samples, num_neg_samples )
+    execStr = '%s %s %s %s %s -numPos %d -numNeg %d -maxFalseAlarmRate 0.45 -featureType LBP -numStages 13 -precalcValBufSize 0 -precalcIdxBufSize 0' % (TRAINCASCADE, data_arg, vector_arg, bg_arg, width_height_arg, num_pos_samples, num_neg_samples )
 
-    print "Execute the following command to start training:"
-    print execStr
+    print("Execute the following command to start training:\n%s" % execStr)
     #opencv_traincascade -data ./out/ -vec ./positive/vecfile.vec -bg ./negative/negative.txt -w 120 -h 60 -numPos 99 -numNeg 5  -featureType LBP -numStages 8
     #opencv_traincascade -data ./out/ -vec ./positive/vecfile.vec -bg ./negative/negative.txt -w 120 -h 60 -numPos 99 -numNeg 5  -featureType LBP -numStages 20
 elif command == "SDFLSDFSDFSDF":