Search

Simple Summary: Identifying bird species is crucial in various bird monitoring tasks. In this study, we implemented a bird species recognition method based on visual features by convolutional neural networks. This paper proposed an attention-guided data enhancement method and a decoupled knowledge distillation model compression method. Using the two methods, we created a novel and efficient lightweight fine-grained bird classification model. The model not only achieved high accuracy in bird classification but was also user-friendly for edge devices such as mobile. Our work may be useful for bird species recognition and bird biodiversity monitoring. Classifying birds accurately is essential for ecological monitoring. In recent years, bird image classification has become an emerging method for bird recognition. However, the bird image classification task needs to face the challenges of high intraclass variance and low inter-class variance among birds, as well as low model efficiency. In this paper, we propose a fine-grained bird classification method based on attention and decoupled knowledge distillation. First of all, we propose an attention-guided data augmentation method. Specifically, the method obtains images of the object's key part regions through attention. It enables the model to learn and distinguish fine features. At the same time, based on the localization–recognition method, the bird category is predicted using the object image with finer features, which reduces the influence of background noise. In addition, we propose a model compression method of decoupled knowledge distillation. We distill the target and nontarget class knowledge separately to eliminate the influence of the target class prediction results on the transfer of the nontarget class knowledge. This approach achieves efficient model compression. With 67% fewer parameters and only 1.2 G of computation, the model proposed in this paper still has a 87.6% success rate, while improving the model inference speed. [ABSTRACT FROM AUTHOR]

Simple Summary: The timely and accurate detection of ewe estrus behavior in precision animal husbandry is an important research topic. The timely detection of estrus ewes in mutton sheep breeding can not only protect the welfare of ewes themselves, but also improve the interests of breeding enterprises. With the continuous increase in the scale of mutton sheep breeding and the gradual intensification of breeding methods, the traditional manual detection methods require high labor intensity, and the contact sensor detection methods will cause stress reaction problems in ewes. In recent years, the rapid development of deep learning technology has brought on new possibilities. We propose a method for the recognition ewe estrus based on a multi-target detection layer neural network. The results show that the method can meet the requirements of timely and accurate detection of ewe estrus behavior in large-scale mutton sheep breeding. There are some problems with estrus detection in ewes in large-scale meat sheep farming: mainly, the manual detection method is labor-intensive and the contact sensor detection method causes stress reactions in ewes. To solve the abovementioned problems, we proposed a multi-objective detection layer neural network-based method for ewe estrus crawling behavior recognition. The approach we proposed has four main parts. Firstly, to address the problem of mismatch between our constructed ewe estrus dataset and the YOLO v3 anchor box size, we propose to obtain a new anchor box size by clustering the ewe estrus dataset using the K-means++ algorithm. Secondly, to address the problem of low model recognition precision caused by small imaging of distant ewes in the dataset, we added a 104 × 104 target detection layer, making the total target detection layer reach four layers, strengthening the model's ability to learn shallow information and improving the model's ability to detect small targets. Then, we added residual units to the residual structure of the model, so that the deep feature information of the model is not easily lost and further fused with the shallow feature information to speed up the training of the model. Finally, we maintain the aspect ratio of the images in the data-loading module of the model to reduce the distortion of the image information and increase the precision of the model. The experimental results show that our proposed model has 98.56% recognition precision, while recall was 98.04%, F1 value was 98%, mAP was 99.78%, FPS was 41 f/s, and model size was 276 M, which can meet the accurate and real-time recognition of ewe estrus behavior in large-scale meat sheep farming. [ABSTRACT FROM AUTHOR]