README:s from sample preparation in Training* directories

register-plate recognition with haar-cascade seems to work well enought, using cascade.xml as rekkari.xml from TrainingWithITO40_positivesfromscaled

Author: paapu88

TrainingWithITO40/README

@@ -0,0 +1,67 @@
+# Rekkari
+Recognition of a numberplate of a car
+
+1) Positive samples are in 
+~/PycharmProjects/Rekkari/TrainingWithITO40/PositivePicturesFromPhone
+python3 ../../picture2rectangle.py 10
+Various resolutions are in directories Rectangle*
+unscaled samples with original resolution of the rectangle
+box are in dir 'NotScaled'
+Negative samples are in dir 'NegativeSamples'
+(they have white ball on top of plates)
+clipped images go to dir 'Rectangle*'
+
+accepted full images go to dir 'NotScaled'
+images with rectangle replaced by ball go to dir 'NegativeSamples'
+
+copy positive samples to positive_samples directory
+mkdir ../positive_samples; cp NotScaled/* ../positive_samples
+
+2) copy negative samples to negative_samples directory
+you can generate more negatives by google picture search by
+python3 ../../get_google_images.py "stad bil gata"
+(remember to manually remove positive figures here)
+You can process files by
+python3 ../../../add_balls.py
+# if the 'r' key is pressed, reset the cropping region
+# if the 'c' key is pressed, accept rectanglecp 
+# if the 'k' key is pressed, keep the original
+# if the 'd' key is pressed, do not write image (it is virtually deleted)
+
+which writes to 'HumanProcessed' directory
+Also copy files in cp HumanProcessed/* ../negative_samples/
+
+3) find ./negative_samples -iname "*.jpg" > negatives.txt
+   cp PositivePicturesFromPhone/NotScaled/* positive_images/
+   find ./positive_samples -iname "*.jpg" > positives.txt
+4)
+create distorted positive samples:
+perl ../opencv-haar-classifier-training/bin/createsamples.pl  positives.txt negatives.txt samples 1000 "opencv_createsamples -maxxangle 1 -maxyangle 0.1 -maxzangle 0.1 -maxidev 30 -w 37 -h 10"
+
+check: opencv_createsamples -w 37 -h 10 -vec ./samples/*vec
+
+5)
+merge positive *.vec files to one vec file
+python2 ~/Dropbox/Apu/mergevec.py -v samples -o positives.vec
+#python2 ../opencv-haar-classifier-training/tools/mergevec.py -v samples -o positives.vec
+check: opencv_createsamples -w 37 -h 10 -vec positives.vec
+
+4) generate vec file of positive samples
+NOT USED
+cp positives.txt  info.txt
+edit info.txt to contain pixel info
+> ./positive_images/sample_IMG_20170307_102910.jpg 1 0 0 80 20
+> ...
+opencv_createsamples -num 36 -info info.txt -w 80 -h 20 -vec positives.vec
+
+6) train:
+check: opencv_createsamples -w 37 -h 10 -vec positives.vec
+rm -f classifier/*
+mkdir classifier
+opencv_traincascade -data classifier -vec positives.vec -bg negatives.txt\
+  -numStages 50 -minHitRate 0.999 -maxFalseAlarmRate 0.5 -numPos 1000 \
+  -numNeg 529 -w 37 -h 10 -mode ALL -precalcValBufSize 512\
+  -precalcIdxBufSize 512
+
+7) in rekkariDetection.py play with parameters
+rekkari_cascade.detectMultiScale(img, 1.1, scale)

TrainingWithITO40_positivesfromscaled/README

@@ -0,0 +1,67 @@
+# Rekkari
+Recognition of a numberplate of a car
+
+1) Positive samples are in 
+~/PycharmProjects/Rekkari/TrainingWithITO40/PositivePicturesFromPhone
+python3 ../../picture2rectangle.py 10
+Various resolutions are in directories Rectangle*
+unscaled samples with original resolution of the rectangle
+box are in dir 'NotScaled'
+Negative samples are in dir 'NegativeSamples'
+(they have white ball on top of plates)
+clipped images go to dir 'Rectangle*'
+
+accepted full images go to dir 'NotScaled'
+images with rectangle replaced by ball go to dir 'NegativeSamples'
+
+copy positive samples to positive_samples directory
+mkdir ../positive_samples; cp NotScaled/* ../positive_samples
+
+2) copy negative samples to negative_samples directory
+you can generate more negatives by google picture search by
+python3 ../../get_google_images.py "stad bil gata"
+(remember to manually remove positive figures here)
+You can process files by
+python3 ../../../add_balls.py
+# if the 'r' key is pressed, reset the cropping region
+# if the 'c' key is pressed, accept rectanglecp 
+# if the 'k' key is pressed, keep the original
+# if the 'd' key is pressed, do not write image (it is virtually deleted)
+
+which writes to 'HumanProcessed' directory
+Also copy files in cp HumanProcessed/* ../negative_samples/
+
+3) find ./negative_samples -iname "*.jpg" > negatives.txt
+   cp PositivePicturesFromPhone/Rectangle-37-10/* positive_images/
+   find ./positive_samples -iname "*.jpg" > positives.txt
+4)
+create distorted positive samples:
+perl ../opencv-haar-classifier-training/bin/createsamples.pl  positives.txt negatives.txt samples 1000 "opencv_createsamples -maxxangle 1 -maxyangle 0.1 -maxzangle 0.1 -maxidev 100 -w 37 -h 10"
+
+check: opencv_createsamples -w 37 -h 10 -vec ./samples/*vec
+
+5)
+merge positive *.vec files to one vec file
+python2 ~/Dropbox/Apu/mergevec.py -v samples -o positives.vec
+#python2 ../opencv-haar-classifier-training/tools/mergevec.py -v samples -o positives.vec
+check: opencv_createsamples -w 37 -h 10 -vec positives.vec
+
+4) generate vec file of positive samples
+NOT USED
+cp positives.txt  info.txt
+edit info.txt to contain pixel info
+> ./positive_images/sample_IMG_20170307_102910.jpg 1 0 0 80 20
+> ...
+opencv_createsamples -num 36 -info info.txt -w 80 -h 20 -vec positives.vec
+
+6) train:
+check: opencv_createsamples -w 37 -h 10 -vec positives.vec
+rm -f classifier/*
+mkdir classifier
+opencv_traincascade -data classifier -vec positives.vec -bg negatives.txt\
+  -numStages 50 -minHitRate 0.999 -maxFalseAlarmRate 0.5 -numPos 1000 \
+  -numNeg 529 -w 37 -h 10 -mode ALL -precalcValBufSize 512\
+  -precalcIdxBufSize 512
+
+7) in rekkariDetection.py play with parameters
+rekkari_cascade.detectMultiScale(img, 1.1, scale)