General Description
Deep learning algorithms are applied in image feature searching to process each region of interest (ROI) and extract a series of high-dimensional feature vectors. The extracted feature vectors are compared against every feature vector in an existing repository to find the most similar ones. To further reduce memory footprint and compute overheads, the long feature vectors in floating-point formats are usually quantized to short feature vectors before feature comparison.
The Ascend AI Processor enables hardware-accelerated PQ short feature search and provides APIs for adding/deleting feature repositories, modifying/deleting feature vectors, and performing feature search. It supports a short feature vector of 32 bytes, a configurable number of top results to return, and the 1:N and N:M search modes. You need to prepare a short feature repository and generate ADC tables of the vectors to be searched. The following figure illustrates the basic workflow.
The following figure shows the search and comparison process.
An ADC table is a 32 KB lookup table, whose content needs to be passed to the corresponding API calls in the format of unsigned characters.
The following table describes the feature search specifications of the
|
Type |
Specification Description |
|---|---|
|
Feature length |
32 bytes (The average memory overhead of each feature is about 36 bytes.) |
|
Repository capacity |
Up to 500 million features (depending on the memory capacity of the board) |
|
Search modes |
1:N mode and N:M mode |
|
Number of repositories |
1:N mode: up to 1 million repositories N:M mode: up to 1 repository |
|
Repository capacity |
1:N mode: up to 1 million features N:M mode: up to 10 million features |
|
Search and comparison performance |
5 billion queries per second (QPS) 10 QPS in 500 million features with less than 1s latency |
|
Number of top results to return |
2–4800 |