Neural-symbolic computing aims at integrating two fundamental paradigms in artificial intelligence: machine learning (that is, the ability to learn from examples) and symbolic knowledge representation and reasoning (that is, the ability to reason from what has been learned). The objective of this thesis is the development of a novel neural-symbolic architecture for object detection in natural images. Specifically, this architecture is based on a neuro-symbolic model called Logic Tensor Networks (LTNs). The base concept behind LTNs is the grounding of a first order logic (FOL) that allows to represent symbolic knowledge as operations between tensors in a neural network; LTNs can be trained through stochastic gradient descent to maximize the satisfiability of the FOL.