Zero Short Learning for wildlife imagery

Abstract

This paper introduces an innovative approach for
object detection from wildlife images using Zero-Shot Learning
(ZSL) with the YOLO-World model. Unlike previous object
detection algorithms, which relied on domain-specific training
data, YOLO-World is optimized for zero-shot object recognition,
thus recognizing a wide range of categories without explicit
training on specific labels. The data for this research have been
taken from a dataset pre-processed and pre-trained, already split
into sets of training and testing, such that the accuracy in the
resulting outcome is more precise. Performance evaluation has
been taken with the help of key parameters such as precision,
recall, F1 score, Intersection over Union (IoU), mean Average
Precision (mAP), and proved the adequacy of the model and
its efficiency in detecting highly accurate wildlife objects. The
experimental results highlight the better performance of ZSL in
the detection of wildlife imagery, with a precision of 0.95 and
recall of 0.92, thus achieving a mAP of 0.93 and F1-score of
0.87. A comparative analysis with existing YOLOv3 and YOLOv5
models also highlights the merits of the proposed approach in
wildlife recognition tasks.