Statistical Boosting: A Preprocessing Technique to Enhance Performance of Machine Learning and Deep Learning Models on Partially Occluded Traffic Signs
Abdul Mannan, Kashif Javed, Serosh Karim Noon · 2020
Computer vision based traffic sign recognition is an active field of research with numerous applications such as (1) autonomous driving (2) driver assistance and (3) automatic inventory management. In some cases the traffic signs are degraded due to aging or some portion is suppressed by another nearby object like a tree or a pole etc. In order to facilitate recognition of such signs using feature extraction or convolutional neural networks based techniques, we propose a preprocessing statistical boosting layer. This scheme helps enhancing visible portions and suppresses the occluded parts which consequently makes extracting/learning invariant features more probable. Since there is no benchmark available for occluded and degraded traffic signs, we develop two datasets containing samples from German Traffic Sign Recognition Benchmark (GTSRB) and self collected/generated images. Experimental results show that with the addition of our proposed statistical boosting layer the performance of famous recognition algorithms can be significantly enhanced.