Informative Evaluation of Images Captured by Camera Traps Based on Production Rules: Invited Paper

Raw images of animals captured by camera traps in the wild often have varying quality due to typical artifacts such as low illumination, cluttered forest background, fast moving objects, difficult weather conditions, seasons, inexpensive camera traps, and so on. A modern solution to the problem of animal recognition is the use of deep learning models, which need to be trained on “good” images. Thus, informative evaluation of such images is a hot topic for research. Moreover, it is possible to automatically select images with unknown animals for manual expert recognition. In this study, we examined the quality of images (brightness, contrast, blur, and weather conditions), as well as the shape and position of the animals relative to the camera trap. Based on the extracted features, we created several production rules for making decisions regarding the informative score of still images collected in Ergaki National Park, Krasnoyarsk region, Russian Federation. This multi-class classification problem was successfully solved by dividing raw images into eleven classes (in terms of subsequent deep recognition): “Normal animal attributes”, “Having several animals”, “Poor shape visibility”, “Poor position visibility”, “Brightness correction”, “Blur correction”, “Weather artifacts”, “Human presence”, “Unrecognized image”, “Unrecoverable image”, and “Object absence”.