Crime Scene Investigation Image Retrieval with Fusion CNN Features Based on Transfer Learning
Ying Liu, Yanan Peng, Daxiang Li, Jiulun Fan, Yun Li · 2018
Crime scene investigation (CSI) image retrieval plays an important role in police case solving. In order to relieve the problem of over-fitting in the training of Convolutional Neural Network (CNN) model due to the limited number of CSI images, this paper presents a CNN feature fusion method based on transfer learning algorithm. Firstly, the pre-trained models VGG-F and VGG-Very Deep Convolutional Network 16 (VGG-VD 16) are fine-tuned, then the features of Fully Connected Layer 7 (FC7) are extracted from the two models respectively, and Principal Component Analysis (PCA) method is used to reduce the features dimensions. Then the features from both models are fused as the final image feature. Tested on CSI image dataset, experimental results show that the proposed algorithm has very good performance.