A Transfer Learning based approach for Pakistani Traffic-sign Recognition; using ConvNets
Zain Nadeem, Abdul Samad, Zulkafil Abbas, Janzaib Massod · 2018
Several effective methods of traffic-sign recognition have been around for a lot of time now, starting with recognition using conventional Image Processing techniques which are very generic and sluggish. However, majority of state-of-the-art detectors are based on Convolutional Neural Networks (CNNs) which have been evidenced to be de facto leader in image processing and computer vision research over the past decade. This has been made possible by datasets being easily available, organized and maintained with German Traffic Sign Recognition Benchmark being of relevance. CNNs require colossal amounts of data to work well; unfortunately, no traffic-sign dataset exists in Pakistan to enable any detector based on CNNs. This paper presents an approach revolving around transfer learning whereby, a model is pre-trained using German Traffic-sign Dataset and is then fine-tuned over Pakistani Dataset: which is collected across Pakistan and amounts to 359 images. Preprocessing and regularization are used to improve overall performance of the model. The fine-tuned model reached training accuracies of around 41% with minimal overfitting. This presents an encouraging outcome as even with a dataset which is comparatively meager, we have achieved a respectable accuracy, something which can be built upon and bettered by boosting number of images collected.