<![CDATA[Wildlife monitoring systems based on camera traps require automated classification methods capable of processing thousands of images with high accuracy to support endangered species conservation. One of the main challenges in animal image classification is the complexity of vegetation backgrounds and dynamic lighting, which are often considered visual distractions for deep learning models. The use of background subtraction methods aims to separate fauna objects from their backgrounds to enhance the model's focus on key animal features. However, the effectiveness of background subtraction on the classification accuracy of Indonesian fauna species still needs further evaluation. This study develops a classification model by evaluating the impact of background subtraction using four segmentation methods (DeepLabV3+, PSPNet, YOLOv11n-seg, and Mask R-CNN) integrated with MobileNetV2 and ResNet-50 architectures. Experimental results on 10,800 images of 9 rare fauna species show that background subtraction significantly reduces accuracy, with a decrease of up to 33% for YOLOv11n-seg and 13-14% for Mask R-CNN. Conversely, models without background subtraction achieved the highest accuracy of 98%. These findings identify that the background in camera trap images is not noise, but vital contextual information that helps the model recognize species.]]>
Copyrights © 2026
