While there are lots of papers about thedetection and recognition of traffic signs, the detection oftheir defects are not well discovered yet. In our paper wediscuss different neural network approaches to find variouserrors on already detected traffic signs. We introduce adata-set of over 4000 items including three frequent errortypes: covered, faded, and scribbled. Two major approachesare investigated: convolutional neural networks to learnthe features of defects, and siamese convolutional neuralnetworks to compare traffic signs with others with knowndistortions. While the former models are known for theirgood performance in object recognition in general, the laternetworks are often used for the detection of defects ofobjects. Neither approach requires information about thetype of the traffic sign itself. We also introduce a techniqueto post-process the confidence values of siamese networks,obtained on different input pairs, to improve accuracy. Thebest results we could achieve was 0.89 F1-score on our data-set
(PDF) Deep Neural Network Models for the Recognition of Traffic Signs Defects. Available from: |
