基于实例分割的白羽肉鸡体质量估测方法

针对白羽肉鸡体质量测量自动化水平低、易造成肉鸡应激的问题，提出一种结合深度学习的非接触式白羽肉鸡体质量估测方法。利用Mask R-CNN和YOLACT(You only look at coefficients) 两种实例分割算法获取白羽肉鸡位置与覆盖掩膜，并进行效果对比；采用自适应掩膜随机提取白羽肉鸡身体部分边缘点，并作为观测点进行椭圆拟合，映射白羽肉鸡背部像素投影面积；通过双变量相关性分析验证白羽肉鸡背部投影面积与体质量间的显著相关性，根据白羽肉鸡背部投影面积与背部像素投影面积的线性比例关系，按照最小二乘原则建立白羽肉鸡背部像素投影面积与体质量间的线性回归模型。试验表明，单只鸡体质量估测中以Mask R-CNN进行特征提取的体质量估测平均准确率为97.23%，以YOLACT进行特征提取的体质量估测平均准确率为97.49%，群鸡场景中体质量估测最低准确率为90.50%。

Breeding White Feather Broiler Weight Estimation Method Based on Instance Segmentation

With the low-automation and stress problem of breeding white feather broiler, a non-contact weight estimation method combined with deep learning was proposed to estimate the weight of breeding white feather broilers quickly and accurately. Mask R-CNN and YOLACT (You only look at coefficients) was used to obtain the target mask and locate the target with position coordinate. The breeding white feather broilers can be completely stripped out from complex background. Then, the edge points of body were extracted for ellipse fitting, and the pixel body area can be obtained. Bivariate correlation analysis was used to show the significant correlation between body weight and body area which was linearly proportional to the pixel body area. The linear regression model between target pixel body area and body weight was established based on the least-square principle. The experimental results showed that the proposed method had a good effect. This method can accurately estimate the body weight of 28-week-old and 48-week-old breeding white feather broilers with different occasion, such as the ideal posture, the head extension, the head turning and partial occlusion. The average accuracy based on Mask R-CNN feature extraction was 97.23%, and the average accuracy based on YOLACT feature extraction was 97.49%. The lowest accuracy for single broiler in the group was 90.50%. The weight of breeding white feather broilers can be estimated quickly and accurately.