Specific Objective Recognition Based on Convolutional Neural Network Transfer Learning

S94 objective recognition not only classifies what category it is, but also detects sub-categories information to confirm the objective is what we want to find. Due to the problem that the number of labeled specific objective images is insufficient, and the subtle different between the specific objective and other objectives in the same category, we proposed an efficient method based on convolutional neural network (CNN) transfer learning. We adopted VGG16 network structure to learn low level visual feature, and the fully-connected layers were redesigned. The method retrained the last fully-connected layer and soft-max classifier, and fine-tuned the model with backpropagation algorithm. The experimental results show that the method makes it possible to use CNN fine-tuning for specific objective recognition with only small scale images, and effectively improves the accuracy of image classification.