Ensemble deep neural networks for domain-specific Image Recognition
Wenbo Li, Chuan Ke · 2016
The dog breeds recognition is the topic of the MSR Image Recognition Challenge (IRC) @ IEEE ICME 2016. In order to sufficiently train the deep neural networks (DNNs) used in this domain-specific image recognition task, we append two high quality public dog datasets and one noisy dog pictures dataset crawled from web to the Clickture-Dog dataset provided by the Challenge. At the preprocessing stage, we enhance the edge of images to emphasize the boundaries of the dogs in image. At the same stage, we adopt the multi-scale training strategy on both the original images and images with edge enhancement. Therefore, we have total four different preprocessing configurations. Finally, we integrate 12 base-DNN sub-models and soft-max posteriors into a committee machine. We implement our recognition algorithm and convert it to Microsoft cloud service. Our team, named as NFS2016, get the third place in this challenge.