A Novel Ensemble Based Reduced Overfitting Model with Convolutional Neural Network for Traffic Sign Recognition System

Abstract

The ELVD (Ensemble-based Lenet VGGNet and DropoutNet) model is used in this paper to examine hypothetical principles and theoretical identification of a real-time image classification and object, tracking, and recognition device running on board a vehicle. Initially, we obtained the dataset from Kaggle. After loading the images, they were converted into 4D tensors and then into a grid. The model has to set the training to 70% training and 30% testing. The ELVD model uses 39,209 32 × 32-pixel color images for preparation, and 12,630 images specifically for research, in the GTSD (German Traffic Sign Detection) dataset. Each picture is a photograph of a traffic sign that corresponds to one of the 43 classes. The picture is a 32 × 32 × 3 sequence of pixel quality values in the RGB color region, defined as pixel values. The image’s class is hidden as a numerical value from 0 to 42. The image collection is somewhat unbalanced, and a few classes are represented significantly better than in the alternative model. The contrast and brightness of the images also differ significantly. The proposed model was created with CNN with Keras and applied with ensemble-based combined LeNet, VGGNet, and DropoutNet pooling layer for tuning the information. The proposed model compares the predicted class with the correct class for all input images and time calculation for predicting different road sign detection images. Underfitting is shown by a standard of low accuracy on the training and testing sets. For a small dataset, the trained model achieved a 98% accuracy level which implied that overfitting the model with the best results on classification accuracy, tested with 15 epochs, resulted in a loss of information of 0.059% and test accuracy of 98%, respectively. Next, the ELVD proposed models trained and validated with different class presents, dataset 2 achieved 93% training accuracy and testing accuracy predicted with 91%. Finally, the ELVD proposed model predicted the test results of unseen class information measured with the 60/km ph, which predicted 99% accuracy. The proposed model predicted noisy as well as unseen multiclass information with fast-moving vehicles. The usage of convolutional layer filter with ensemble-based VGGNet, DropouNet, and LeNet trained and predicted with a high classification accuracy of more than 99% combined ELVD model with fastest time calculation also the high accuracy prediction of selected image dataset labels that enables these models to be used in real-time applications. The ELVD model was also compared with other traditional models of VGGNet, LeNet, and DropoutNet; its detection time outperformed the other models, and it achieved a 98% detection label set of information. In the ELVD model, closure to various human abilities on a related responsibility differs from 97.3% to 99.5%; consequently, the ELVD model performs better than an average human.