如何在 keras 中为损失函数实现 ssim？

Question

I need SSIM as a loss function in my network, but my network has 2 outputs. I need to use SSIM for first output and cross-entropy for the next. The loss function is a combination of them. However, I need to have a higher SSIM and lower cross-entropy , so I think the combination of them isn't true. Another problem is that I could not find an implementation of SSIM in keras.

Tensorflow has tf.image.ssim , but it accepts the image and I do not think I can use it in loss function, right? Could you please tell me what should I do? I am a beginner in keras and deep learning and I do not know how can I make SSIM as a custom loss function in keras.

Answer 1

Have a look here for SSIM loss in Keras. You can write a custom loss function and create SSIM loss for one prediction and cross-entropy for another. You can return a weighted sum of the two losses as the final loss.

Answer 2

other choice would be ssim_loss = 1 - tf.reduce_mean(tf.image.ssim(target, output, max_val=self.max_val))

then combine_loss = mae (or mse) + ssim_loss In this way, you are minimizing both of them.