利用深度學習於商業豬舍中自動偵測和時序分割母豬生產仔豬事件

Abstract

肉豬是全球畜牧業的重要基石，亦在台灣畜牧業擁有很高的經濟價值。母豬分娩階段是商業豬舍中特別重要的時期，降低仔豬出生後的死亡率仍是持續存在的挑戰。傳統監測方法高度依賴人工觀察，不僅勞力密集、判定主觀。在現代集約化生產系統的趨勢下愈發不符合成本效益。本研究的主要目標是利用電腦視覺與深度學習，開發一套能夠在商業分娩欄位中自動偵測個別仔豬出生事件的方法。本研究自兩座商業豬舍共21個分娩欄位收集了467部仔豬生產影片。該資料集的類別數量極端不平衡，其中3.6%的影像「生產」，96.4%的影像則為「休息」。本研究最終開發了一套自動偵測母豬生產仔豬模型。該模型採用兩階段架構，包含空間特徵提取器與時序分類器（MS-TCN++）。本研究針對兩種特徵提取器：MobileNetV3-Large與DINOV2-S進行測試，並比較了預訓練(pretrain)、微調(fine-tuning)及特徵融合(fusion)等不同方法，同時採用焦點損失函數(focal loss)來解決類別不平衡問題。經微調並結合特徵融合的MobileNetV3-Large達到最佳模型表現，F1@0.1為78.18%、F1@0.25為67.27%。採用特徵融合的預訓練DINOV2-S則達到F1@0.1為70.48%、F1@0.25為53.33%的模型表現。在資料集無法對特定領域進行微調的情況下，展現出一種替代方案。由PDDM預測結果所計算出的分娩統計數據，包括，胎數、總分娩時長、平均生產間隔及平均生產持續時間，其數值均與實際數值呈現高度一致性。本研究為精準畜牧業奠定基礎，期望透過自動量化高精度的時序分娩統計資訊，輔助豬農即早介入母豬異常分娩。

Pork production represents a cornerstone of global livestock agriculture and holds high economic value among animal husbandry industry in Taiwan. The farrowing phase is a particularly critical period in commercial pig production. Decreasing the mortality rate of pre-weaning piglets remains a persistent challenge. Traditional monitoring methods rely heavily on manual observation, which is labor-intensive, subjective, and increasingly impractical given the rising trend in modern intensive production systems. The primary objective of this research was to develop an automated method for detecting individual piglet delivery events within commercial farrowing crates using computer vision and deep learning. A dataset comprising 467 piglet delivery videos from 21 farrowing crates across two commercial pig farms was collected. The dataset exhibited extreme class imbalance, with 3.6% of the frames labeled as "delivering" and 96.4% of the frames labeled as "resting." A Piglet Delivery Detection Model (PDDM) was developed using a two-stage architecture consisting of a spatial feature extractor followed by a temporal classifier (MS-TCN++). Two feature extractors were evaluated: MobileNetV3-Large and DINOV2-S. Different strategies including pretrained, fine-tuned, and feature fusion approaches were compared, with focal loss employed to address the class imbalance problem. The fine-tuned MobileNetV3-Large with feature fusion achieved the best performance with F1@0.1 of 78.18% and F1@0.25 of 67.27%. The pretrained DINOV2-S with feature fusion achieved F1@0.1 of 70.48% and F1@0.25 of 53.33%, presenting as a compelling alternative when domain-specific fine-tuning is not possible. Farrowing statistics derived from PDDM predictions, including litter size, total farrowing duration, mean delivery interval, and mean delivery duration, demonstrated strong agreement with ground truth values. This research lays a foundation for precision livestock farming, aiming to assist pig farmers in early intervention of abnormal sow farrowing through automated quantification of high-fidelity temporal farrowing statistics.