Interpreting motion in the natural world presents a major challenge for visual systems. Natural scenes vary enormously in structure, luminance and contrast, all parameters known to modulate the response of biological motion detectors. Nevertheless, many animals overcome this challenge and adopt visually guided behaviour for which the accurate estimation of self-motion and image velocity is required. It is generally accepted that Reichardt correlator-like computations underlie local motion detection in insects. Reichardt correlators, however, generate ambiguous estimates of velocity, because they are sensitive to several additional image parameters, such as those mentioned above. How does the visual system generate accurate estimates of apparent image velocity when the elements underlying local motion detection produce ambiguous velocity signals? This thesis investigates the neural processing of image velocity. I performed sharp electrode intracellular recordings from identified motion sensitive neurons in the lobula plate of the hoverfly, Eristalis tenax. A series of natural and artificial images were used to investigate the processing of a vast range of scenes. I show that the horizontal system (HS) neurons have a remarkable capacity to estimate image velocity reliably for vastly different natural scenes. This property is at odds with the HS neurons' responses to experimenter-defined stimuli. I reveal several activity dependent features of the neural response that may reconcile the ability to accurately encode the velocity of natural images with the mechanisms underlying motion processing. Images that were initially weak neural drivers have long latencies, with responses continuing to increase in magnitude over several hundred milliseconds. Images that were initially strong neural drivers, reached peak responses more rapidly followed by significant reductions in response over longer time scales. Despite being different in sign and time course, these two activity dependent changes in response act as near-ideal normalisers for images that would otherwise produce highly variable response magnitudes. By analysing the time course of neural response and manipulating image contrast, I show that this property is likely to emerge from a combination of static and dynamic non-linarities. When image contrast is reduced, thus reducing the range of input signals to local motion detectors, the essential non-linearity of the Reichardt correlator model provides a good prediction of global responses. Thus, suggesting an important role for non-linear mechanisms being recruited by high contrast local features in the robust encoding of natural scenes. Finally, I use an experimental paradigm that reduces the influence of spatial integration and thus enables the analysis of responses equivalent to the outputs of individual local motion sensitive elements presynaptic to the HS neuron. I show evidence for an adaptive gain reduction that affects the sensitivity of individual motion detector responses to subsequent features. This gain reduction is facilitated by local neighbouring motion stimulation and is thus, well suited to take advantage of the predictable nature of natural scenes.