Classification of Low Light Vehicles Using Transfer Learning and CNN
R.Divyakanthi1, G.Navyasrivenu 2, D.Udayasrisailakshmi3 &G.Shanmukhsai4
1 Assistant Professor
[2-4] B. Tech Student, LIET
[1,2,3,4] Electronics & Communication Engineering, Lendi Institute of Engineering and Technology, Vizianagaram
-------------------------------------------------------***---------------------------------------------------------
Abstract – The main Aim of the project is to enhance the low light images and extract features using the Transfer learning approach of Convolution Neural Network (CNN) integrating with enhancement technique for increasing the vehicle recognition rate. The images are enhanced using Low Light Image Enhancement (LIME) Method CNN is used for Classification of comp vehicle dataset to evaluate the proposed Model. This paper focuses on the classification of vehicles using Convolutional Neural Network (CNN) which is a class of deep learning neural network. This work makes use of transfer learning using the pre-trained networks to extract powerful and informative features and apply that to the classification task. In the proposed method, the pre-trained networks are trained on two vehicle datasets consisting of real-time images. The classifier performance along with the performance metrics such as accuracy, precision, false discovery rate, recall rate, and false negative rate is estimated for the following pre-trained networks Google Net. The classification model is implemented on the standard vehicle dataset and also on a created dataset. The model is further used for the detection of the different vehicles using Regions with a Convolutional Neural Networks (RCNN) object detector on a smaller dataset. This paper focuses on finding the perfect network suitable for the classification problems which have only a limited amount of non-labelled data. The model makes use of limited pre-processing and achieves greater accuracy on continuous is training of the networks on the vehicle images.
Key Words: Convolutional Neural Network, low light Enhancement, Google Net [DAG Network with properties], Matlab, Image Data Store, Deep learning tool box, precision, Accuracy, Validation, pre-trained networks, dataset.
