Optimasi Arsitektur You Only Look Once Version 7 (YOLOv7) untuk Deteksi Perilaku Agresif pada Ayam Pedaging

Abstract

Saat ini, Computer Vision telah banyak diterapkan di sektor peternakan untuk membantu pemantauan hewan ternak. Perilaku agresif merupakan salah satu permasalahan umum pada unggas berkoloni seperti ayam pedaging yang dapat berpengaruh terhadap efisiensi produktivitas dan kesejahteraan ternak. Penelitian sebelumnya menggunakan YOLOv4 berhasil membangun model deteksi perilaku agresif pada ayam pedaging namun performa deteksi model masih cukup rendah. Penelitian ini berfokus pada peningkatan presisi model dalam mendeteksi perilaku agresif pada ayam pedaging dengan mengimplementasikan arsitektur deteksi objek You Only Look Once versi ketujuh (YOLOv7) serta menerapkan penyesuaian label untuk meningkatkan kualitas data. Hasil pelatihan menggunakan arsitektur YOLOv7 menghasilkan nilai mAP (Mean Average Precision) sebesar 24,6% meningkatkan performa model YOLOv4 sebelumnya sebesar 17,1% dalam deteksi perilaku agresif. Selanjutnya, model YOLOv7 dengan penyesuaian label pada data mendapatkan nilai mAP yang lebih baik yaitu sebesar 53,7% yang menunjukkan bahwa penyesuaian label berpengaruh terhadap kinerja model.

Currently, Computer Vision has been widely applied in the livestock sector to help monitor livestock. Aggressive behavior is one of the common problems in colonial poultry such as broilers which can affect productivity efficiency and livestock welfare. Previous studies using YOLOv4 have succeeded in building a model for detecting aggressive behavior in broilers, but the detection performance of the model is still quite low. This research focuses on increasing the precision of the model in detecting aggressive behavior in broilers by implementing the seventh version of the You Only Look Once object detection architecture (YOLOv7) and applying label adjustments to improve data quality. The results of training using the YOLOv7 architecture yielded a mAP (Mean Average Precision) value of 24.6%, increasing the performance of the previous YOLOv4 model by 17,1% in the detection of aggressive behavior. Furthermore, the YOLOv7 model with label adjustments on the data obtained a better mAP value of 53.7% which indicated that label adjustments had an effect on model performance.