Masters Thesis

Car Model Classification Using Deep Learning

Considerably deep convolutional neural networks have contributed greatly to recent advances in image recognition. the emergence of architecture like GoogLeNet with Inception Module have allowed neural networks to gain high performance with low computational cost. It is possible to have substantial gains in classification tasks but this directly equates to increased parameter count and computational efficiency trade-offs which Inception architecture addresses. in addition, architectures such as VGG-19 with their considerable depth and small 3 × 3 convolution filters allow us to achieve very high accuracies. the task of car model classification is a challenging, fine-grained task and cars can have broad difference between different makes but at the same time subtle differences in comparison to cars within the same make. in addition, the pose of the car image samples in the dataset as well as the setting the images are captured with can make this task even more challenging. Architectures that are deep and have many layers are best suited for learning such hierarchies. to this end, in this paper, we propose a new model that relies on fusion of VGG and GoogLeNet layers with trained weights on ImageNet dataset allowing us to classify car images at a high rate of accuracy.


In Collection: