علي فرضي پور

In recent years, self-driving cars have been one of the areas of interest for researchers. The advantages of these cars increase the safety of the car and its passengers, which leads to a significant reduction in accidents. Accurate recognition of traffic signs is one of the essentials of car safety. For this, we need to extract the features of the received images and recognize the type of sign using these features. We turn to various deep learning methods to learn image features. One of the most widely used methods is the use of convolutional networks, which have many advantages over neural networks. One of the characteristics of these networks was to pay attention to a specific location of the image. In other words, convolutional network filters extract the relationships between the pixels of a location in the image at each step. With the introduction of the attention mechanism and transformers and their success in the field of natural language processing, researchers turned to using them in machine vision challenges. The use of transformers showed that their mechanism extracts the relationships between each pixel of the image and the rest of the pixels. In other words, vision transformers pay attention to the image in general. In this work, we have presented a new model by combining both methods to take advantage of both of them. This model is based on transformer and convolution. In order to improve the model in the layer close to the input of the model in order to extract the features more simply and better, as well as to reduce the calculations, we have used 1x1 convolutions with different number of filters to extract the features and then 3x3 convolution to find the relationships between these features. Also, in order to prevent memorizing data instead of learning it, we have used a combined data augmentation method. To eva‎luate the introduced model, we have compared it with 5-7 other well-known models in 3 different datasets . We have been able to achieve 99.66% accuracy in the first dataset and 99.8% in the second dataset with the proposed model, which is more than other works. Also, in the dataset of 61 parts of the third challenge, we achieved better results than other models. These results indicate that the proposed model, in addition to higher accuracy, also performs better in challenging conditions.

ترنسفورمر , يادگيري عميق , تشخيص اشيا , خودروي خودران

transformer , deep learning , object recognition , self-driving car

Ali Farzipour

Dr. Shahriar Baradaran Shokouhi