deep-anpr
deep-anpr copied to clipboard
Change in detect.py can improve tolerance to noise
I changed detect.py in order to improve efficiency, the detection in real images is very sensitive to noise, so instead of detecting the highest probability it returns the most voted plate:
plates = []
for letter_prob in letter_probs:plates.append(letter_probs_to_code(letter_prob))
mostcommonplate =collections.Counter(plates).most_common(1)[0][0]
yield (numpy.max(mins, axis=0).flatten(),
numpy.min(maxs, axis=0).flatten(),
numpy.max(present_probs),
mostcommonplate
)
I have obtained much higher accuracy results using this approach and the detect algorithm is more robust.