Object recognition and localization are in focus of the research field in image processing. Although we recently developed some effective methods, we have no exact solution for the problem yet. The most promising way is deep learning neural network supported by high computational performance. This method is generally inspired by biological systems and mimics the process of network connections in brain. I primarily focus on optical recognition of vehicles and first I’m going to observe it as a part of machine learning and give a historical traverse. The task contains several parts such as constructing the structure in line with the complexity of the problem, and training on the image dataset (each image has a label). I’m going to introduce the common cost functions (cross entropy) with its regularization terms (L1,L2), the gradient descent and backpropagation algorithm. It is important to speak about bias variance problem and possible solutions (for example dropout) with debug methods. Due to the lack of computational performance I’m going to rely on already trained networks and analyze their structure: it is common to keep some earlier layers fixed and only tune the higher level region of the network. This thesis takes emphasis on mathematical approaches and introduces the most popular deep learning software packages. In the end I’m going to evaluate the neural network image classification performance on a test set.