仿真假體視覺下神經網絡算法的應用研究

趙瑛 李琦 王冬暉 于愛萍 谷宇

摘? 要： 盡管已有多種圖像處理策略被應用到視覺假體的仿真研究中并提高了被試的識別表現，但在植入電極數量有限的情況下，如何保證盲人獲得足夠的拓撲信息是視覺假體仍需解決的問題。在此背景下，本文將兩種神經網絡算法應用到仿真假體視覺中對圖像進行前景目標提取和像素化處理，首先利用圖像分割數據集訓練一個U?net網絡得到前景提取后的結果，將其像素化之后與提取前的原圖配對，再利用配對后的數據集訓練一個Pix2Pix網絡從而實現了將彩色圖像“翻譯”為像素化圖像的目標。實驗結果表明，與傳統圖像處理算法相比U?net網絡具有更準確的目標提取效果，且經Pix2pix網絡“翻譯”后的圖像也與標簽圖像更相似，有助于提高假體佩戴者的識別準確率。

關鍵詞： 仿真假體視覺; 神經網絡算法; 前景目標提取; 像素化處理; 數據集訓練; 圖像配對

中圖分類號： TN911.34?34; TP391.4? ? ? ? ? ? ?文獻標識碼： A? ? ? ? ? ? ? ? ? ? ? 文章編號： 1004?373X（2020）04?0164?03

Application of neural networks algorithm in artificial prosthesis vision

ZHAO Ying1， LI Qi1， WANG Donghui1， YU Aiping1， GU Yu1，2

（1. Inner Mongolia Key Laboratory of Pattern Recognition and Intelligent Image Processing， School of Information Engineering， Inner Mongolia University of Science and Technology， Baotou 014010， China; 2. School of Computer Engineering and Science， Shanghai University， Shanghai 200444， China）

Abstract： A variety of image processing strategies have been applied to the simulation study of visual prosthesis and improved the recognition performance of the subjects， but how to ensure that blind people get enough topological information with limited electrodes is still a problem to be solved in visual prosthesis. On this background， two kinds of neural network algorithm are applied to the artificial prosthesis vision to perform foreground object extraction and pixel processing for images. A U?net network is trained with the image segmentation dataset to acquire the extracted results of foreground， which are matched with the original image after its pixel processing. A Pix2pix network was trained with the matched datasets to achieve the goal of “translating” color images into pixelated images. The experimental results showed that in comparison with the traditional image processing algorithm， the U?net network has more accurate object extraction effect， and the image translated by the Pix2Pix network is more similar to the ground truth. It is helpful to improve the recognition accuracy of prosthesis wearers.

Keywords： artificial prosthesis vision; neural networks algorithm; foreground object extraction; pixel processing; dataset training; image matching

人類對外界信息的獲取約有80%來自于視覺，視力殘疾或失明對人的正常生命活動造成的影響是災難性的。對于由視網膜色素變性（Retinal Pigmentosa）和老年性黃斑變性（Age?related Macular Degeneration）造成視力殘疾或失明的患者[1]，他們病變的視網膜組織中仍然有近80%的內層神經細胞和近30%的神經節細胞的形態和功能處于正常狀態，并保持一定的功能性連接，視覺假體可應用于這類患者的視覺恢復和重建過程中[2]。通過在他們的視網膜、視覺皮層或視神經等位置植入微電極并對其施加合適的電刺激，可使患者感知到光幻視點。……