[9p-Z10-1~11] 【CS.3】 3.3 情報フォトニクス・画像工学と4.4 Information Photonicsのコードシェアセッション

▲ [9p-Z10-6] Deep Learning for Single-Pixel Imaging Without Normalization and Image Output

〇(PC)Masaki Yasugi^1,2、Yasuhiro Mizutani³、Takeshi Yasui⁴、Hirotsugu Yamamoto^1,2 (1.Utsunomiya Univ.、2.JST, ACCEL、3.Osaka Univ.、4.Tokushima Univ.)

キーワード：single-pixel imaging, deep learning

Single-pixel imaging is an interesting method to reconstruct an image by use of a photodetector and modulated illuminations. Although a large number of illumination patterns enable us to reconstruct a fine image, it takes so much time to obtain the image. Recently, researchers have tried to drastically reduce the number of illumination patterns by using deep learning, on numerical simulations and actual experiments. The reconstructed data itself is obtained as array data calculated by the correlation between detected intensity and illumination. However, the data is normalized and negative values seems to be changed to zero before the output to reconstructed image. In this study, we examine the effect of the type (array or image) and normalization of reconstructed data on its restoration through the neural network, by numerical simulation. Our neural network can restore fine image from reconstructed data without normalization and image output, particularly in the case of many masks.