Enhanced YOLOv8 for Real-Time Stray Animal Detection in Challenging Visibility Conditions

Stray animal detection is crucial in urban management and wildlife conservation. However, under challenging visibility conditions, the low contrast and blurry edges of the targets and their tendency to blend into the background lead to incorrect identification. To address these challenges, we propose an improved YOLOv8 model that utilizes the Adam optimizer to improve the generalization capability. We introduce an attention mechanism to optimize small target detection and incorporate Graph Convolutional Networks (GCN) to enhance spatial feature representation. Additionally, deconvolution techniques are employed to align contextual information, while structure-guided image enhancement improves the quality of low-light images. Experimental results demonstrate that our method outperforms baseline models in terms of accuracy, recall, and precision on the Stray Cats and Dogs Datasets and the African Wildlife Datasets, while maintaining real-time detection capabilities.