Classify 42 vehicle logos with different CNN models
We explore different models with Keras to do this job. Our dataset is about 25,000 and divided about 8:2 into training and validation dataset. Each image is resized to 64x64, so we change most models to fit our data which is originally designed for 224x224 image size.
- cifar10 A simple CNN model with just 4 Conv layers and relative few filters of kernel, achieves about 96% accuracy in val.
- AlexNet Classic AlexNet model with smaller first-layer kernel size and stride, achieves about 97.4% accuracy in val.
- VggNet Classic VggNet16 model without last block, achieves state-of-the-art about 98.5% accuracy in val.
- ResNet Classic ResNet model of new improved version, but performs not very well in this task with only 92% accuracy in val. We are trying to find results.
All the above networks are equiped with Batch Normalization, which is a effective method that helps the neural network converge quickly and achieve relative good results in just few epochs.