An Advanced Chicken Face Detection Network Based on GAN and MAE

SIMPLE SUMMARY: Chicken face detection is a fundamental task for accurate poultry management. Achieving satisfactory chicken face detection is necessary to implement downstream tasks, such as day-age detection, behavior recognition, and health monitoring. Nonetheless, the image dataset of the chicken face is small-scale, and there are few related studies. Moreover, chicken heads and features are smaller than other livestock, making recognition tricky. Inspired by these significances and obstacles, this paper proposes a chicken face detection network with an augmentation module. Based on the YOLOv4 backbone, our model achieved 0.91 F1, 0.84 mAP, and 37 FPS, far surpassing the two-stage RCNN and EfficientDet baselines. This model can be applied to an actual chicken coop, and its performance is adequate to conduct downstream tasks. ABSTRACT: Achieving high-accuracy chicken face detection is a significant breakthrough for smart poultry agriculture in large-scale farming and precision management. However, the current dataset of chicken faces based on accurate data is scarce, detection models possess low accuracy and slow speed, and the related detection algorithm is ineffective for small object detection. To tackle these problems, an object detection network based on GAN-MAE (generative adversarial network-masked autoencoders) data augmentation is proposed in this paper for detecting chickens of different ages. First, the images were generated using GAN and MAE to augment the dataset. Afterward, CSPDarknet53 was used as the backbone network to enhance the receptive field in the object detection network to detect different sizes of objects in the same image. The [Formula: see text] feature map output was added to three feature map outputs of this paper, thus changing the feature map output of eightfold downsampling to fourfold downsampling, which provided smaller object features for subsequent feature fusion. Secondly, the feature fusion module was improved based on the idea of dense connection. Then the module achieved feature reuse so that the YOLO head classifier could combine features from different levels of feature layers to capture greater classification and detection results. Ultimately, the comparison experiments’ outcomes showed that the mAP (mean average Precision) of the suggested method was up to 0.84, which was 29.2% higher than other networks’, and the detection speed was the same, up to 37 frames per second. Better detection accuracy can be obtained while meeting the actual scenario detection requirements. Additionally, an end-to-end web system was designed to apply the algorithm to practical applications.