研究目的
Investigating the problem of automatically determining what type of shoe left an impression found at a crime scene, focusing on cross-domain image matching using deep features and similarity metrics.
研究成果
The proposed multi-channel normalized cross-correlation (MCNCC) significantly improves cross-domain image matching performance, achieving state-of-the-art results in shoeprint retrieval and other applications. Discriminative training further enhances performance, demonstrating the effectiveness of deep features and learned similarity metrics for forensic and other cross-domain tasks.
研究不足
The approach assumes decorrelated channels for computational efficiency, which may not hold in all cases. Performance can degrade with high background clutter or severe occlusions. Limited training data for learning models may lead to overfitting. The method requires pre-aligned images or exhaustive search over alignments, increasing computational cost.
1：Experimental Design and Method Selection:
The study uses a Siamese network architecture for cross-domain image matching, employing multi-channel normalized cross-correlation (MCNCC) as the similarity measure. It involves diagnostic experiments and cross-domain matching tasks with various datasets.
2：Sample Selection and Data Sources:
Datasets include an internal Israeli dataset of shoeprints (387 test impressions and 137 crime scene prints), the FID-300 benchmark (1175 test impressions and 300 crime scene prints), the CMP Facade Database (606 facade images and segmentation labels), and a Google Maps dataset (2194 pairs of aerial photos and map images).
3：List of Experimental Equipment and Materials:
Pre-trained CNN models (ResNet-50, GoogLeNet, DeepVGG-16) are used for feature extraction. Software tools include MATLAB for CCA and custom code for implementation.
4：Experimental Procedures and Operational Workflow:
Features are extracted from images using pre-trained CNNs. Similarity is computed using MCNCC with local per-channel normalization. For shoeprints, searches over translations and rotations are performed. Learning involves discriminative training with hinge loss and backpropagation.
5：Data Analysis Methods:
Performance is evaluated using precision-recall curves, cumulative match characteristic (CMC), and mean average precision. Statistical analysis includes ablation studies and visualization of feature influences.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
