Improved Multi-Scale Object Detection Accuracy on Battery Chicken Coop with YOLO and Features Pyramid Network

This research aims to increase object detection accuracy in chickens in a battery cage environment with various obstacles, such as feeding places, drinking places, and iron cages. The algorithm used is You Only Look Once (YOLO), which was modified by adding a Feature Pyramid Network (FPN) and several additional layers to increase detection accuracy. Images were annotated using Label Studio, and data augmentation was performed using Albumentations with a total of 29,060 annotations. The dataset was taken from livestock pens in the North Sulawesi region, Indonesia. Adjustment of the spatial size of the tensor is carried out using a forward mechanism using the F.interpolate function with the nearest neighbor method.The research results show that the YOLO modification with FPN produces a mean Average Precision (mAP) value of 0.88567, better than standard YOLO with mAP of 0.87616. Experiments were carried out on various numbers of epochs (25, 50, 75, and 100) with the best results at the 100th epoch, achieving a precision of 0.85876, recall of 0.80472, mAP50 of 0.88567, and mAP50-90 of 0.59254. In addition, a comparison is made between the proposed method and the YOLOv8 standard. The results show that YOLOv8 + FPN has precision advantages in all evaluation metrics over standard YOLOv8, especially in the values mAP50 and mAP50-90, with an increase in mAP of 4.83%.This research proves that adding FPN and additional layers to YOLOv8 effectively increases object detection accuracy in a chicken coop environment with multi-scale conditions and object obstacles. These findings significantly contribute to the development of computer vision-based object detection technology.