Detection of Objects Through Hierarchical Version of Fast Region-Based Convolutional Neural Networks

EasyChair Preprint no. 5434

Abstract

In object detection there is high degree of skewedness for objects' visual separability. It is difficult towards distinguishing certain categories demanding dedicated classification. The training for the deep convolutional neural networks (CNNs) is performed through N-way classifiers. Considerable work needs to be done for leveraging structures in hierarchical category. We present here hierarchical fast region-based CNNs (Hrch Fast RCNNs) where deep CNNs are embedded considering hierarchy as categorical. The easy classes are separated through classifiers in coarse category. The difficult classes are classified by classifier in fine category. The training in Hrch Fast R-CNN is achieved by initial training of the components which follows fine-tuning globally using multiple group discriminant analysis. The regularization is done using consistency in coarse category. For large-scale recognition tasks, scalability is done considering conditional execution of classifiers in fine category and compression in layer parameters. Using CIFAR100 datasets as benchmark we obtain good results. We build four different Hrch Fast R-CNN where standard CNNs top-1 error are reduced significantly.

Keyphrases: Classification, CNN, object detection, recognition, Scalability

@Booklet{EasyChair:5434,
  author = {Arindam Chaudhuri},
  title = {Detection of Objects Through Hierarchical Version of Fast Region-Based Convolutional Neural Networks},
  howpublished = {EasyChair Preprint no. 5434},

  year = {EasyChair, 2021}}