The Camouflage Machine: Optimizing protective coloration using deep learning with genetic algorithms

Analysing biological colour patterns from digital images: An introduction to the current toolbox

Understanding the numerous roles that colouration serves in the natural world has remained a central focus in many evolutionary and ecological studies. However, to accurately characterise and then compare colours or patterns among individuals or species has been historically challenging. In recent years, there have been a myriad of new resources developed that allow researchers to characterise biological colours and patterns, specifically from digital imagery. However, each resource has its own strengths and weaknesses, answers a specific question and requires a detailed understanding of how it functions to be used properly. These nuances can make navigating this emerging field rather difficult. Herein, we evaluate several new techniques for analysing biological colouration, with a specific focus on digital images. First, we introduce fundamental background knowledge about light and perception to be considered when designing and implementing a study of colouration. We then show how numerous modifications can be made to images to ensure consistent formatting prior to analysis. After, we describe many of the new image analysis approaches and their respective functions, highlighting the type of research questions that they can address. We demonstrate how these various techniques can be brought together to examine novel research questions and test specific hypotheses. Finally, we outline potential future directions in colour pattern studies. Our goal is to provide a starting point and pathway for researchers wanting to study biological colour patterns from digital imagery.

CamoEvo: An open access toolbox for artificial camouflage evolution experiments

Camouflage research has long shaped our understanding of evolution by natural selection, and elucidating the mechanisms by which camouflage operates remains a key question in visual ecology. However, the vast diversity of color patterns found in animals and their backgrounds, combined with the scope for complex interactions with receiver vision, presents a fundamental challenge for investigating optimal camouflage strategies. Genetic algorithms (GAs) have provided a potential method for accounting for these interactions, but with limited accessibility. Here, we present CamoEvo, an open-access toolbox for investigating camouflage pattern optimization by using tailored GAs, animal and egg maculation theory, and artificial predation experiments. This system allows for camouflage evolution within the span of just 10-30 generations (∼1-2 min per generation), producing patterns that are both significantly harder to detect and that are optimized to their background. CamoEvo was built in ImageJ to allow for integration with an array of existing open access camouflage analysis tools. We provide guides for editing and adjusting the predation experiment and GA as well as an example experiment. The speed and flexibility of this toolbox makes it adaptable for a wide range of computer-based phenotype optimization experiments.

Generalist camouflage can be more successful than microhabitat specialisation in natural environments

BACKGROUND

Crypsis by background-matching is a critical form of anti-predator defence for animals exposed to visual predators, but achieving effective camouflage in patchy and variable natural environments is not straightforward. To cope with heterogeneous backgrounds, animals could either specialise on particular microhabitat patches, appearing cryptic in some areas but mismatching others, or adopt a compromise strategy, providing partial matching across different patch types. Existing studies have tested the effectiveness of compromise strategies in only a limited set of circumstances, primarily with small targets varying in pattern, and usually in screen-based tasks. Here, we measured the detection risk associated with different background-matching strategies for relatively large targets, with human observers searching for them in natural scenes, and focusing on colour. Model prey were designed to either 'specialise' on the colour of common microhabitat patches, or 'generalise' by matching the average colour of the whole visual scenes.

RESULTS

In both the field and an equivalent online computer-based search task, targets adopting the generalist strategy were more successful in evading detection than those matching microhabitat patches. This advantage occurred because, across all possible locations in these experiments, targets were typically viewed against a patchwork of different microhabitat areas; the putatively generalist targets were thus more similar on average to their various immediate surroundings than were the specialists.

CONCLUSIONS

Demonstrating close agreement between the results of field and online search experiments provides useful validation of online citizen science methods commonly used to test principles of camouflage, at least for human observers. In finding a survival benefit to matching the average colour of the visual scenes in our chosen environment, our results highlight the importance of relative scales in determining optimal camouflage strategies, and suggest how compromise coloration can succeed in nature.