Analisis Robustness Model MobileNetV2 dalam Mengatasi Occlusion untuk Identifikasi Wajah Sapi

Abstract

Dalam upaya meningkatkan efisiensi dan keberlanjutan industri peternakan melalui integrasi Teknologi Informasi dan Komunikasi (TIK), sistem identifikasi ternak yang akurat dan otomatis menjadi komponen penting untuk menggantikan metode konvensional. Identifikasi sapi berbasis biometrik wajah menghadapi tantangan akibat oklusi visual di lingkungan kandang yang dinamis. Penelitian ini mengkaji pengaruh strategi pelatihan Synthetic Occlusion untuk meningkatkan ketahanan (robustness) model MobileNetV2 dalam mengidentifikasi wajah sapi teroklusi. Dataset terdiri atas 1.414 citra Region of Interest (RoI) yang diperoleh dari rekaman video sapi. Strategi ini menginjeksi synthetic occlusion ke dalam proses pelatihan tanpa mengubah arsitektur model. Model baseline tanpa augmentasi oklusi digunakan sebagai pembanding untuk mengukur efektivitas pendekatan yang diusulkan. Evaluasi melalui stress testing pada tiga tingkat oklusi menunjukkan bahwa strategi ini mampu menahan degradasi performa secara signifikan. Dibandingkan model baseline yang mengalami penurunan akurasi hingga 0,65 pada oklusi berat, model proposed mempertahankan akurasi sebesar 0,8481 pada oklusi sedang dan 0,75 pada oklusi berat, dengan akurasi kondisi ideal mencapai 0,9816. Temuan ini menunjukkan bahwa synthetic occlusion dapat meningkatkan ketahanan model tanpa memerlukan pengumpulan data teroklusi secara manual. Sistem diimplementasikan dalam Progressive Web App (PWA) untuk penggunaan di lapangan.

In an effort to improve the efficiency and sustainability of the livestock industry through the integration of Information and Communication Technology (ICT), accurate and automated livestock identification systems have become essential to replace conventional identification methods. Facial biometric-based cattle identification, however, remains challenging due to visual occlusion commonly encountered in dynamic farm environments. This study investigates the effect of a Synthetic Occlusion training strategy on improving the robustness of a MobileNetV2 model for occluded cattle face identification. The dataset consists of 1,414 Region of Interest (RoI) facial images extracted from cattle video recordings. The proposed strategy injects synthetic occlusions into the training pipeline without modifying the underlying network architecture. A baseline model trained using standard augmentation without synthetic occlusion was employed as a benchmark to evaluate the effectiveness of the proposed approach. Stress testing under three occlusion levels demonstrated that the proposed strategy substantially reduced performance degradation. While the baseline model's accuracy decreased to 0.65 under severe occlusion, the proposed model maintained accuracies of 0.8481 under moderate occlusion and 0.75 under severe occlusion, while achieving 0.9816 under ideal conditions. These findings indicate that synthetic occlusion effectively enhances model robustness without requiring the manual collection of naturally occluded training data. Finally, the trained model was deployed as a Progressive Web App (PWA) to support practical field implementation in livestock farming.