PDIC: AN UNSUPERVISED ENHANCEMENT FRAMEWORK FOR DETECTION-AWARE NIGHT-TIME WILDLIFE IMAGING

Abstract

Wildlife detection in the night time is a critical computer vision problem that has applications in ecological monitoring, road safety surveillance, and conservation analysis. In real deployments, camera imaging systems are often found to be photon limited, which means that they are oper ating in an environment where the density of photons is low, and the noise level in the sensor as well as the object boundaries, motion blur, and infrared effects are also large.In real deploy ments, the density of photons, the noise level of the sensor, the boundaries of the objects, the motion blur, and the infrared effects are often large, causing the objects to be difficult to detect in a reliable manner. While the YOLO family of detectors can make fast inference in daylight images, they fail to perform well in dark, low-contrast night images with cluttered backgrounds and underexposed animals [1, 2, 37]. This thesis introduces a new unsupervised image-enhancement pipeline called Progres sive Dual-Branch Illumination-Contrast (PDIC) specifically developed as an enhancement pre-processing step prior to a fixed YOLOv5 detector for night-time wildlife detection. In PDIC, the architecture of the detector is not changed, rather the visible evidence presented to the de tector is enhanced. The framework is founded on the idea of decomposing image information based on the Retinex, and consists of two interconnected branches: an Illumination Compen sation Branch for estimating and normalizing spatial illumination distribution, and a Contrast Enhancement Branch for enhancing the gradient of structures and mid-tone details and con trolling the amplification of noise. The enhanced reflectance and illumination components are combined again in a learnable fusion module and progressive refinement allows the framework to process heavily underexposed images more than once [9, 8, 7, 4]. Training is completely unsupervised, that is, no paired low light and normal light images of wildlife are needed. The goal is to minimize smoothness of the illumination, gradient uni formity, exposure loss, color constancy loss, and total variation loss. These losses are chosen to maintain degradation of edges and contours relevant to the detectors and not just for human vi sual optimisation. The framework is applicable in practical camera trap use cases, where paired