研究目的
To propose a coarse-to-fine method for fast indexing with Deep Convolutional Neural Network (DCNN) and Local Geometrical Constraint Model to improve accuracy and efficiency in large-scale image retrieval.
研究成果
The proposed coarse-to-fine method effectively combines DCNN, PCA, LSH for efficient coarse retrieval and a novel Local Geometrical Constraint Model for high-precision fine retrieval. It handles non-rigid deformations and outliers robustly, outperforming state-of-the-art methods in experiments on near-duplicate image datasets. Future work could focus on reducing computational complexity and extending to more general image retrieval tasks.
研究不足
The computational complexity of the fine-grained method is O(N^3) due to solving linear systems with Gram matrices, which may be prohibitive for very large N. The method assumes the availability of initial putative correspondences and may be sensitive to parameter settings (e.g., λ, η, τ, γ). It is tested primarily on near-duplicate images and may not generalize well to highly diverse image sets.
1：Experimental Design and Method Selection:
The methodology involves a coarse-to-fine schema. Coarse-grained retrieval uses DCNN for feature extraction, PCA for dimension reduction, and LSH for hashing to generate binary codes for efficient similarity computation via Hamming distance. Fine-grained retrieval employs a Local Geometrical Constraint Model based on a Bayesian framework with latent variables for inlier/outlier classification, solved using the Expectation Maximization (EM) algorithm in a reproducing kernel Hilbert space (RKHS) with Tikhonov and local regularization.
2：Sample Selection and Data Sources:
The dataset is a mix of the California-ND dataset and images from JD.com, totaling 100,000 images. For fine-grained evaluation, 120 images from 12 classes in California-ND are used, generating 7260 image pairs. Initial matches are obtained using SIFT features from the VLFEAT toolbox.
3：List of Experimental Equipment and Materials:
A laptop with a 2.5-GHz Intel Core CPU and 8-GB memory is used for experiments. Software includes the Caffe framework for DCNN feature extraction and VLFEAT for SIFT matching.
4：5-GHz Intel Core CPU and 8-GB memory is used for experiments. Software includes the Caffe framework for DCNN feature extraction and VLFEAT for SIFT matching.
Experimental Procedures and Operational Workflow:
4. Experimental Procedures and Operational Workflow: In coarse-grained step, features are extracted with DCNN, reduced with PCA, and hashed with LSH to produce binary codes for Hamming distance-based ranking. Top 120 ranked images are selected as candidates. In fine-grained step, the Local Geometrical Constraint Model is applied to candidate pairs to compute preserved matches as similarity, using EM for transformation estimation and outlier removal.
5：Data Analysis Methods:
Performance is evaluated using precision and recall metrics for both retrieval and matching. Precision is the ratio of correct retrievals or matches to total, and recall is the ratio of correct retrievals or matches to total possible correct ones. Comparisons are made with state-of-the-art methods like RANSAC, ICF, and GS.
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
