以影像辨識為基礎之仔豬異常偵測系統

在台灣初生仔豬被壓死及體弱發育不良仔豬是養豬產業最容易被疏忽的損失之一，初生仔豬死亡首要因素是被母豬壓死，其原因為母豬生產後十分虛弱導致無力避開仔豬及母豬躺下也是最容易壓死小豬的情況。母豬哺乳時也時常發生仔豬未進食的情況，原因可能仔豬體型較為瘦弱及遭排擠。養豬戶須不斷的巡視豬場，經常在發現仔豬異常時，以搶救不及。因此本研究使用影像辨識技術建置一套「以影像辨識為基礎之仔豬異常偵測系統」，利用影像前處理技術，再導入卷積神經網路進行學習，本研究之實驗場域為南部某養豬場，在豬舍裝設攝影機，並錄製5個月的豬場影像，每日以12小時做記錄，幀寬度為600像素，幀高度為500像素，建立影像資料庫。實驗結果顯示母豬哺乳辨識率為94%、遊走仔豬辨識率為91%、初生仔豬辨識率為85%，可降低養豬養豬戶的人力負擔及減少養豬業者的損失。

關鍵字

影像辨識；影像前處理；仔豬；卷積神經網路

並列摘要

In Taiwan, the death of newborn piglets due to crushing, as well the frail and stunted piglets, is among the most neglected losses in the pig industry. The primary cause of the death of newborn piglets is due to crushing from sows. The reason for this is that the sows are so weak after they give birth to piglets that they were unable to keep away from the piglets. In addition, it is most likely for sows to crush piglets to death when they lie down. Piglets often cannot eat when sows are suckling. The reason for this may be that piglets are relatively thin and they are pushed aside. Pig farmers must constantly inspect the pig farm so that they can take measures when piglets are found to be abnormal. Therefore, this study used the image recognition technology to build “An Abnormality Detection System for Piglets Based on Image Recognition”, which uses the image preprocessing technology, and then imports the convolution neural network for learning. The experimental field of this study is located in a pig farm in the south. In this study, a video camera was installed in the pig house and a video of the pig farm for 5 months was recorded. It recorded video for 12 hours every day, with the frame width of 600 pixels and the frame height of 500 pixels, so as to create the image database. The experimental results showed that the identification rate of sow suckling was 94%, the identification rate of moving piglets was 91%, and the identification rate of newborn piglets was 85%, which could reduce the manpower burden of pig farmers and reduce the losses of practitioners in the pig industry.