This paper presents and compares the image classification methods based on MLPs and CNNs. Training data is 500,000 pictures of 10 different objects. The first architecture to be used is MLPs network that contains 3,853,298 weights, the second architecture is CNNs with 528,054 weights. This paper proposes several methods and architectures network to avoid overfitting phenomenon and increases the accuracy of modeling approximately 80%. Besides on it, the paper also presents and compares time training of models using CPUs, and combining CPUs with GPUs.