Image Hash Model Training with Center Collision
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image retrieval methods require large storage space and high computational workload due to the extraction of thousands of image features, leading to decreased retrieval efficiency and accuracy.
Innovation Solution
An image processing method using a trained image hash model that projects similar images to the same center point in a space, allowing for hash code retrieval instead of feature comparison, thereby reducing complexity and improving efficiency and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If thousands of image features are extracted to improve retrieval accuracy, then retrieval accuracy is improved, but storage space requirements and computational workload increase significantly
Solution Approach 1:
The patent transforms image features from thousands of continuous dimensions to fixed-length binary hash codes (e.g., 64-bit or 128-bit). This parameter transformation dramatically reduces storage requirements and enables efficient Hamming distance-based comparison, while the learned hash function preserves semantic similarity information to maintain retrieval accuracy
Solution Approach 2:
The patent extracts only the essential similarity information from high-dimensional image features and encodes it into compact binary hash codes. By taking out and retaining only the critical features needed for similarity judgment, the system achieves both storage efficiency and retrieval accuracy
2Measurement precision
If thousands of image features are extracted to improve retrieval accuracy, then retrieval accuracy is improved, but retrieval speed decreases due to heavy computational workload
Solution Approach 1:
The patent replaces complex continuous feature comparison with binary hash code comparison using Hamming distance. This substitution transforms the computational mechanism from heavy floating-point operations to simple bit-wise XOR operations, dramatically improving retrieval speed while maintaining accuracy through learned hash functions that preserve semantic relationships
3Device complexity
If traditional hash methods are used to reduce complexity, then storage and computation are simplified, but retrieval accuracy decreases due to loss of similarity information
Solution Approach 1:
The patent performs preliminary learning training to teach the hash function how to map similar images to the same hash buckets. By pre-training the model with similarity information before actual retrieval operations, the system prepares the hash function to preserve semantic relationships, ensuring both efficiency and accuracy are achieved
Data Source
Figure 1A~1B
Figure 2
Figure 3~4
AI summary
An image processing method and apparatus, and a computer device. The method comprises: obtaining, according to a convolutional neural network, corresponding feature points of various training images in a feature embedding space, and a feature center point of respective images of the types, to which the training images belong, of a second number; performing network training by means of central collision to obtain a target feature center point corresponding to an image of each type; and then, mapping the obtained various feature points and various target feature center points to a Hamming space to obtain a first number of Hash codes and a second number of Hash center points, in which case, network training can still be performed by means of the central collision to obtain an image Hash model. By means of learning the similarities between the Hash codes of the various training images and the similarities between the feature points of the training images and the corresponding central points, global distribution of images is not only learned, but the network training time is also greatly shortened, and the learning efficiency and the precision of the image Hash codes are improved.