Adaptive image processing embedding to make the ecological tasks of deep learning more robust on camera traps images

Abstract

Camera traps serve as a valuable tool for wildlife monitoring, generating a vast collection of images for ecologists to conduct ecological investigations, such as species identification and population estimation. However, the sheer volume of images poses a challenge, and the integration of deep learning into automated ecological investigation tasks remains complex, particularly when dealing with low-quality images in long-term monitoring programs. Existing approaches often struggle to strike a balance between image enhancement and deep learning for ecological tasks, thereby overlooking crucial information contained within low-quality images. This research introduces a pioneering adaptive image processing module (AIP) that seamlessly incorporates image processing into camera trap ecological tasks, elevating the performance of wildlife monitoring activities. Specifically, a differentiable image processing (DIP) module is presented to enhance low-quality images, with its parameters predicted by a Non-local based parameter predictor (NLPP). Additionally, an end-to-end approach based on hybrid data containing both original and synthetic data is proposed, encompassing adaptive image processing methods and downstream tasks for camera traps, adaptable to various scenarios. This approach effectively reduces the manual labor and time required for professional image processing. When applied to real-world camera trap images and synthetic image datasets, our method achieves an accuracy of 92.26% and 86.65% in classifying wildlife, respectively, demonstrating its robustness. By outperforming alternative methods under harsh conditions, the application of the adaptive image processing module instills greater confidence in deep learning applications within complex environments.