研究目的
To develop a learnable, DNN-based descriptor optimized for retinal image registration and compare its performance to hand-crafted descriptors.
研究成果
DNN-based feature descriptors trained on retinal image patches perform better than binary descriptors (BRISK, FREAK) but worse than floating-point hand-crafted ones (SIFT, SURF, KAZE). CNN+FC architectures with LRN and training on hard samples showed better results. Future work should focus on improved dataset augmentation and inclusion of general image patches to enhance performance.
研究不足
Dataset limitations may affect performance, as not all geometric distortions in real images were simulated during augmentation. The learned descriptors did not outperform top hand-crafted ones like SIFT and SURF, indicating room for improvement in training and dataset construction.
1：Experimental Design and Method Selection:
Used a Siamese network architecture for training DNN-based feature descriptors, with contrastive loss function and ADAM optimization. Compared various DNN structures (FCN, CNN+FC with LRN or BN) and hand-crafted descriptors (SIFT, SURF, BRISK, FREAK, KAZE).
2：Sample Selection and Data Sources:
Compiled a training dataset from nine online retinal image datasets (Chase DB, Diaret DB, DTSET1, DTSET2, HRF-base, Messidor1, Messidor2, Messidor3, RODREP), totaling 3153 initial images, augmented to 11279 images. Used FIRE dataset for evaluation with ground truth point pairs.
3：List of Experimental Equipment and Materials:
No specific equipment mentioned; utilized software tools like MATLAB toolbox for KAZE and VLFeat for SIFT.
4：Experimental Procedures and Operational Workflow:
Keypoints detected using multiple detectors (FAST, Harris, MinEigen, Hessian, BRISK, SURF, SIFT, KAZE), clustered with k-means, and patches (101x101 pixels) cropped. Augmentations included flipping, rotation (up to 10 degrees), projective distortions, and scaling. Patches resized to 64x64 for training. Generated 17 million training samples (half similar, half non-similar pairs). Trained DNN models with and without hard samples, evaluated on FIRE dataset using Rank-1 metric.
5：Data Analysis Methods:
Computed Euclidean distances between descriptors, sorted matches, and calculated Rank-1 performance (proportion of correct matches in top position).
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
