Image generation method, image generation apparatus, and image generation program

Abstract

Embodiments are directed to accurately measuring a distance between a “true probability distribution: q” and a “probability distribution determined from a model of a generator: p” by D(x,y) of cGANs, so that a generated image may be made closer to a true image. A method of generating an image by using a conditional generative adversarial network constituted by two neural networks which are a generator and a discriminator, in which the discriminator outputs a result obtained from an arithmetic operation using a model of the following equation:f(x, y; θ):=f1 (x, y; θ)+f2(x; θ)=yTV ϕθΦ(x)+ψθΨ (ϕθΦ (x))

1 - 6 . (canceled) 
     
     
         7 . A model training device comprising:
 at least one memory; and   at least one processor configured to:   provide at least one of (1) first data generated by a generator or (2) training data to a first neural network included in a discriminator,
 calculate a first scalar value based on a product between an output of the first neural network and condition information, 
 calculate a loss based on the first scalar value, and 
 update, based on the loss, at least one of the discriminator or the generator. 
   
     
     
         8 . The device according to  claim 7 , wherein the condition information is represented by a one-hot vector. 
     
     
         9 . The device according to  claim 8 , wherein the one-hot vector includes category information. 
     
     
         10 . The device according to  claim 7 , wherein the at least one processor is further configured to:
 calculate a second scalar value by providing the output of the first neural network to a second neural network included in the discriminator, and   calculate the loss based on the first scalar value and the second scalar value.   
     
     
         11 . The device according to  claim 10 , wherein the at least one processor is further configured to calculate the loss by adding the first scalar value and the second scalar value. 
     
     
         12 . The device according to  claim 7 , wherein the first data generated by the generator is image data. 
     
     
         13 . The device according to  claim 7 , wherein the at least one processor is further configured to input noise to the generator to generate the first data. 
     
     
         14 . A data generation device comprising:
 at least one memory; and   at least one processor configured to:
 generate data by inputting noise to a generator trained by the device according to  claim 7 . 
   
     
     
         15 . A model training method comprising:
 providing, by at least one processor, at least one of (1) first data generated by a generator or (2) training data to a first neural network included in a discriminator,   calculating, by the at least one processor, a first scalar value based on a product between an output of the first neural network and condition information,   calculating, by the at least one processor, a loss based on the first scalar value, and   updating, by the at least one processor, at least one of the discriminator or the generator based on the loss.   
     
     
         16 . The method according to  claim 15 , wherein the condition information is represented by a one-hot vector. 
     
     
         17 . The method according to  claim 16 , wherein the one-hot vector includes category information. 
     
     
         18 . The method according to  claim 15 , further comprising:
 calculating, by the at least one processor, a second scalar value by providing the output of the first neural network to a second neural network included in the discriminator,   wherein calculating the loss includes calculating the loss based on the first scalar value and the second scalar value.   
     
     
         19 . The method according to  claim 18 , wherein calculating the loss includes calculating the loss by adding the first scalar value and the second scalar value. 
     
     
         20 . The method according to  claim 15 , wherein the first data generated by the generator is image data. 
     
     
         21 . The method according to  claim 15 , further comprising:
 inputting, by the at least one processor, noise to the generator to generate the first data.   
     
     
         22 . A data generating method comprising:
 generating, by at least one processor, data by inputting noise to a generator trained by the method according to  claim 15 .   
     
     
         23 . A non-transitory computer readable medium storing a program configured to cause at least one computer to execute a method comprising:
 providing at least one of (1) data generated by a generator or (2) training data to a first neural network included in a discriminator,   calculating a first scalar value based on a product between an output of the first neural network and condition information,   calculating a loss based on the first scalar value, and   updating at least one of the discriminator or the generator, based on the loss.   
     
     
         24 . The non-transitory computer readable medium according to  claim 23 , wherein the condition information is represented by a one-hot vector. 
     
     
         25 . The non-transitory computer readable medium according to  claim 24 , wherein the one-hot vector includes category information. 
     
     
         26 . The non-transitory computer readable medium according to  claim 23 , wherein the method further comprises:
 calculating a second scalar value by providing the output of the first neural network to a second neural network included in the discriminator,   wherein calculating the loss includes calculating the loss based on the first scalar value and the second scalar value.