US 12,368,952 B2
Auto focus (AF) method and system for electro-hydraulic (EH) lens with adjustable focus, and an electronic device
Jun Liu, Shandong Province (CN); Hengyu Li, Shanghai (CN); Yunyun Jiang, Shanghai (CN); Jingyi Liu, Shanghai (CN); Yueying Wang, Shanghai (CN); Shaorong Xie, Shanghai (CN); and Jun Luo, Shanghai (CN)
Assigned to Jining University, Qufu (CN); and Shanghai University, Shanghai (CN)
Filed by Jining University, Shandong Province (CN); and Shanghai University, Shanghai (CN)
Filed on Dec. 8, 2023, as Appl. No. 18/534,020.
Claims priority of application No. 202311094832.X (CN), filed on Aug. 28, 2023.
Prior Publication US 2025/0080842 A1, Mar. 6, 2025
Int. Cl. H04N 23/67 (2023.01); G02B 3/14 (2006.01); G06V 10/22 (2022.01); G06V 10/25 (2022.01); G06V 10/82 (2022.01); H04N 23/617 (2023.01)
CPC H04N 23/673 (2023.01) [G02B 3/14 (2013.01); G06V 10/22 (2022.01); G06V 10/25 (2022.01); G06V 10/82 (2022.01); H04N 23/617 (2023.01); G06V 2201/07 (2022.01)] 13 Claims
OG exemplary drawing
 
1. An auto focus (AF) method for an electro-hydraulic (EH) lens with an adjustable focus, comprising:
determining a state space, an action space, and a reward function of a reinforcement learning (RL) method, wherein the state space comprises at least two parameters, respectively being image definition and a camera focal length; the action space comprises one parameter, being a focusing current value of the EH lens with an adjustable focus; and the reward function is a function designed according to an image difference before and after an action;
obtaining a target image acquired by an image sensor, automatically selecting a focusing target area from the target image by using a computer vision technology, and calculating image definition and a camera focal length of the focusing target area;
inputting a current state into a policy network in a deep neural network (DNN) architecture to obtain a current initial action, adding noise to the current initial action to obtain a current composite action, and determining a next state according to the current composite action, wherein the current state refers to image definition and a camera focal length of a current focusing target area; the current initial action refers to a current focusing current value; and the next state refers to image definition and a camera focal length of a next focusing target area;
calculating a current reward according to the current state, the next state, the current composite action, and the reward function, and storing the current state, the next state, the current composite action, and the current reward as a set of sample data in an experience pool;
using sample data in the experience pool as training data of the DNN architecture when there are M sets of the sample data in the experience pool, and obtaining a trained AF policy with reference to a deep deterministic policy gradient (DDPG) algorithm and a single hill climbing optimization (HCO) algorithm; and
deploying the trained AF policy to a to-be-focused EH lens with an adjustable focus, so that the to-be-focused EH lens with an adjustable focus is capable of automatically adjusting a focal point in a real-time environment.