Recent research has shown that deep learning techniques outperform traditional steganography and steganalysis methods, which has contributed in several researches to propose different types of increasingly complex and larger convolutional neural networks (CNNs) to detect steganographic images, which aims to outperform the state of the arts most of the time in a 1%-2%. This paper presents a data preprocessing and distribution strategy that improves accuracy and convergence during training. The strategy implements a bifurcation of spatial rich model (SRM) filters and DCT filters, which are a set on one branch as trainable and on the other untrainable, followed by three blocks of residual convolutions and an excitation layer. The proposed strategy improves the accuracy of CNNs applied to steganalysis by 2%-15% while preserving the stability.