70 / 2025-04-15 16:50:18

Intelligent Identification of Abnormal Feces and the Floor Fouling Degree in Growing and Finishing Pens Based on Computer Vision

Pig, diarrhea, pen fouling, inspection robot, computer vision

3. 智慧牧场与动物福利 > （3）智慧养殖的传感器、智能感知和机器人

Timely detection of abnormal feces and pen fouling is crucial for effective pig health monitoring, precise control of the pig house environment, and efficient manure removal. Traditional manual on-site observation methods were often time-consuming, labor-intensive, subjective, and lack timeliness. To address these challenges, this study proposed an intelligent identification method for detecting abnormal feces and assessing the degree of floor fouling in pig pens, leveraging computer vision technology. The proposed approach consists of three main steps: acquiring the original overhead image, performing pig segmentation and masking, and dividing the image into sub-images for classification. Pig instance segmentation in the original image is achieved using the YOLOv11 model. The masked original image is then divided into six sub-images for further classification. A seven-point (A to G) image classification standard was developed based on factors such as the presence of slats, feces, urine, and abnormal feces. A dataset of 26,670 sub-images was constructed through manual labeling, including 8,099 A-type (solid floor, dry, no feces), 3,214 B-type (solid floor, urine only), 1,099 C-type (solid floor, normal feces only), 1,973 D-type (solid floor, both urine and normal feces), 3,391 E-type (solid floor, abnormal feces), 7,613 F-type (slatted floor, no abnormal feces), and 1,281 G-type (slatted floor, abnormal feces). The dataset was randomly split into training and validation sets at a 9:1 ratio and fine-tuned using the Swin Transformer model. The results demonstrate that the YOLOv11 algorithm achieved a mean Average Precision (mAP@50) of 0.8214 for pig segmentation on the validation set. Swin Transformer achieved a classification accuracy of 86.38% for sub-image classification. Notably, the recognition accuracy for abnormal feces on solid and slatted floors was 85.43% and 86.60%, respectively, with corresponding recall rates of 88.20% and 85.16%. This study provides robust algorithms and theoretical support for mobile inspections of pig health and pen cleanliness, offering significant potential for improving pig farming management.