Abstarct: Digital images play an important role in our daily lives and can be used for a variety of applications such as object recognition, biometric authentication, and video surveillance. Noise is an undesirable signal that causes an accidental change in the amount of color or brightness in an image. Sometimes, images get infected with noise due to many reasons such as defects in camera sensors and transmission in a noisy channel. Denoising is a basic need to enhance and retrieve hidden and valuable details in an image; therefore, an efficient method is needed to image denoising. Deep convolution neural networks have attracted a lot of attention in the field of image denoising. However, there are two problems: first, it is very difficult to train a convolutional network to denoising, and second, that most of the deep parts of the network become saturated and a vanishing gradient occurs. In this dissertation, a new network called AutoencoderNet is designed. Instead of increasing the depth to extract the feature, we used increasing the bandwidth. In this method, we combine the two networks as an upper and lower network, which increases the width instead of increasing the depth. In the upper grid, we use an autoencoder to retrieve the input image at the output, and in the lower grid, we use the dilated convolutions to extract more features to remove noise. Also, to prevent the data distribution from changing and not being dependent on the initial value, batch normalization is used, and to facilitate training and prevent vanishing gradient, residual learning is used in the design of the AutoencoderNet network. Experimental results show that the PSNR value has improved by 1.27 dB for gray images and 1.22 dB for color images compared to previous methods.