I have a signal matrix 256x256. which contains signal peaks( complex values) in some rows other are zero. I am adding noise to particular rows which contains values and try to reconstruct the same signal using Convolutional autoencoder.(not using any standard model) . I am creating my own model with 5 layers to do but while testing with unseen data, in reconstruction part sometimes i am getting peaks with noise or getting noise only in whole matrix. (noise vary 0DB to -10DB). can any one please suggest something how can i improve my results.