Posture Key Frame Recognition and Analysis for Weight Estimation of Yellow-feathered Chickens

Abstract:

Body weight is a key indicator to evaluate the growth condition of poultry. However, the variation of poultry posture will affect the accuracy of weight estimation. SE-ResNet18+fLoss network was proposed to detect the posture key frames of floor-reared yellow-feathered chickens. The attention mechanism SE module and residual structure were integrated. And the Focal Loss was added to solve the problem of sample imbalance. In addition, the Gradient-weighted Class Activation Mapping was introduced to explain the rationality of the end classification rule. The dataset was constructed by 4295 images of yellow-feathered chickens. The F1-score of the SE-ResNet18+fLoss model on the test set for the chicken situations recognition of six classes: standing, bowing head, spreading wing, grooming feather, sitting and occlusion were 94.34%, 91.98%, 76.92%, 93.75%, 100% and 93.68%, respectively. Towards the detection of key posture frames on chickens, the accuracy, recall, F1-score and detection speed were 97.38%, 97.22% 97.26% and 19.84f/s, respectively. And the detection accuracy, recall and F1-score were better than those of ResNet18, MobileNet V2 and SE-ResNet18 networks. The study ensured real-time performance while improving the accuracy of key posture frame recognition, which provided technical support for accurate estimation of poultry weight.