CNN regression for Denoising problem

[mau]

Hi all.
I'm trying to train a CNN for denoising of images.

Layers scheme is:

data -> convolution (k output / feature maps) -> activation (sigmoid / tanh) -> convolution (1 output / feature map) -> euclidean loss

where feature map size is equal to input size, with appropriate Pad and stride = 1. K convolution and activation can be repeated many times.

Parameters are chosen with random search algorithm. 

Last layer of convolution product 1 feature maps which represent output of net, so image reconstruction.

My doubt is in last layer: i think filter of last layer of convolution are not able to product appropriate output only using a small filter matrix. Error is too high.
I tried with fully connected last layer, but error doesnt improve.

Any suggestion? 

thanks