Multiclass animal detection in high altitude drone images using LCS-enhanced YOLOv8 with SAHI

The fact that small creatures can be detected in a photograph of wildlife is important in monitoring an ecosystem, assessing the diversity of species or in formulating strategies to protect these creatures. Nevertheless, small creatures are typically a failure case of deep learning techniques since images are not detailed, animals are mixed, or the environment is too chaotic to make any detection reliable. The paper presents a more robust configuration that relies on Linear Contrast Stretching (LCS), YOLOv8, and Slicing Aided Hyper Inference (SAHI) to enhance the performance of small animal detection on drone shots. LCS then changes how the light and contrast of the images are perceived before processing to make animals easier to distinguish between them and their environment. SAHI, conversely, enhances the recognition of small objects by dividing huge images into intersecting parts. This aids in facilitating the process of seeing far away or partly concealed animals. Wildlife image datasets are tested with experiments demonstrating that the methods are great in comparison with a typical YOLOv8 model. Rather than uninteresting YOLOv8, the new LCS-YOLOv8-SAHI configuration achieved an average Precision (mAP) of 0.502 at 0.50 - 65 times higher than the 0.304 of YOLOv8. To make things worse, mAP 0.50:0.95 increased by 54 points, going between 0.14 and 0.216. These figures support the strength and competence of the mixed model in identification of small animals in photographs. With clever application of computer vision and deep learning gimmicks, such an approach provides a legitimate chance to enhance automatic wildlife tracking, to expand conservation efforts, and to advance ecology research.