采用改进RetinaNet的笼养肉鸽繁育期个体检测模型

实现繁育期精准个体检测是提高集约养殖环境下肉鸽繁育效率和精准管控效果的有效手段,其中小目标鸽蛋及粘连乳鸽的精准检测是关键。该研究提出了一种基于改进RetinaNet的目标检测模型,以RetinaNet网络为基础框架,将ResNet50特征提取网络与特征金字塔网络（Feature Pyramid Networks,FPN）结合,增加特征金字塔网络中特征检测尺度,提升对图像中遮挡鸽蛋与粘连乳鸽的检测精度;在分类和回归子网络前引入卷积注意力模块（Convolutional Block Attention Module,CBAM）,提升对小目标检测的精度。试验结果表明,该研究提出的模型对于笼养肉鸽个体检测的平均精度均值（mean Average Precision,mAP）达到80.89%,相比SSD、YOLOv3、YOLOv4、YOLOv5s、YOLOv5m和原始RetinaNet模型提高了18.66、29.15、19.92、21.69、18.99与15.45个百分点;对成鸽、乳鸽与鸽蛋检测的平均精度（Average Precision,AP）分别为95.88%,79.51%和67.29%,相对原始RetinaNet模型提高了2.16、21.74和22.48个百分点,在保证成鸽精准检测的基础上,显著提升了对复杂环境下存在局部遮挡的小目标鸽蛋以及粘连乳鸽的检测精度,为实现集约化养殖环境下肉鸽繁育周期个体检测和精准管控提供有效支持。

Individual detection model for caged meat pigeons during breeding period based on improved RetinaNet

The pigeon industry is emerging after the "three birds" in the special poultry breeding. Pigeon breeding has also been well developed over more than 30 years in the animal husbandry industry in China. In this study, an improved RetinaNet model was proposed for the individual detection of caged pigeons during the breeding period. An accurate and rapid individual detection in the cage was realized to improve the breeding efficiency. The original image dataset was first collected in the Golden Green Pigeon Breeding Base in Xingning, Guangdong Province, China. The images were then enhanced by the vertically flipped, adding noise, and brightness. As such, the training and validation datasets were expanded to five times the original images. A total of 5 420 pigeon images were selected to be labeled, where 5 190 images were set as the training and validation set, and 230 as the test set. The model was improved using the RetinaNet. The Convolutional Block Attention Module (CBAM) was introduced before the classification and regression sub-networks. As such, the information of the feature map was enhanced for the better model. Two backbone networks (ResNet50 and ResNext50) were selected to compare to the Feature Pyramid Networks (FPN) with different layers on the individual detection of pigeons in the breeding period. 11 experiments were then performed on the same training, validation, and test set under the operating system of Ubuntu18.04 and graphics card: NVIDIA TITAN RTX. The commonly-used SSD and YOLOv3 in the one-stage object detection were selected to compare with the RetinaNet model framework in this case. Eight experiments were also carried out to optimize the model using the same dataset and training parameters. The results showed that the mAP of the original RetinaNet model structure was 65.44%, which was 3.21, 13.7, 4.47, 6.24, and 3.54 percentage points higher than that of SSD, YOLOv3, YOLOv4, YOLOv5s, and YOLOv5m, respectively. The overall detection effect of the original RetinaNet model was better than that of SSD and YOLOv3. Moreover, the number of FPN layers increased the detection scale, thereby effectively improving the recall of small objects. The missed detection of pigeon eggs was also reduced in the actual breeding environment using the original RetinaNet. Meanwhile, the CBAM was introduced before the classification and regression sub-network, in order to promote the detection effect on the adhesion pigeons. The improved RetinaNet model presented an average accuracy of 95.88%, 79.51%, and 67.29% on the test set of pigeons, squabs, and pigeon eggs, respectively, which were 2.16, 21.74, and 22.48 percentage points higher than the original RetinaNet model. There was also much more improvement in the average precision of adhesive squab and small pigeon eggs. Consequently, the improved model can also be expected to present the best performance. The real-time monitoring of individual changes can be achieved in the pigeon cage. The finding can provide the technical support for precision breeding, in order to timely detect and adjust the improper behavior of production management.