The textural components of natural images driving the responses of motion-sensitive interneurons

When an animal moves through an environment, images of the surroundings move across it's eyes. This image motion can be evaluated by the animal to gather information relevant for behavioral control, such as in visual navigation, distance estimation or collision avoidance. In the fly, image motion is processed by motion-sensitive neurons, such as the lobula plate tangential cells (LPTCs), i.e. wide-field neurons in the third visual neuropile. The response of LPTCs to moving visual stimuli depends on the contrast and spatial structure of features in the scene. This sensitivity to the textural properties of the images that flow across the retina might therefore provide the fly with potentially valuable information about the environment. Whereas studies on fly motion vision classically use simplified visual stimuli (e.g frequency modulated grating patterns), it is not per se clear, which of the textural components of stimuli with more complex image statistics, such as natural scenes, drive LPTC response. In the current study we aim to identify and characterize textural components of natural scenes that predominantly affect the neural respsonse traces. Textural components are extracted from a dataset of high dynamic range panoramic images using a machine learning approach. This approach allows us to construct a filter bank representing the textural constituents of natural scenes specific for different LPTC receptive field characteristics. This filter bank is subsequently used to assess which of these constituents of natural scenes are predominantly represented in the neuronal responses of a subset of simulated LPTCs. By characterizing the filters based on their power, we are able to predict LPTC response in terms of amplitude and modulation depth. Although the relative contributions of the extracted textural components to the original image is in line with the spatial frequency distribution typically found in natural scenes (approximately 1/f), the mean amplitude and modulation depth of simulated LPTC responses to specific textural components do not follow this rule: textural components with a lower relative occurrence may induce stronger LPTC responses than more predominant image features. Hence, LPTCs respond preferentially to a characteristic range of image features