王海燕,江烨皓,黎煊,马云龙,刘小磊.基于弱监督数据集的猪只图像实例分割[J].农业机械学报,2023,54(10):255-265.
WANG Haiyan,JIANG Yehao,LI Xuan,MA Yunlong,LIU Xiaolei.Pig Image Instance Segmentation Based on Weakly Supervised Dataset[J].Transactions of the Chinese Society for Agricultural Machinery,2023,54(10):255-265.
|
摘要点击次数: 193
全文下载次数: 72
|
| 基于弱监督数据集的猪只图像实例分割 [下载全文] |
| Pig Image Instance Segmentation Based on Weakly Supervised Dataset [Download Pdf][in English] |
| 投稿时间：2023-03-12 |
| DOI：10.6041/j.issn.1000-1298.2023.10.025 |
| 中文关键词: 猪只 弱监督实例分割 空间注意力机制 involution算子 |
| 基金项目:国家重点研发计划项目(2022YFD1601903)、湖北省科技重大专项(2022ABA002)、华中农业大学-中国农业科学院深圳农业基因组研究所合作基金项目（SZYJY2022034）和中央高校基本科研业务费专项资金项目（2662022XXYJ009） |
|
| 中文摘要:在智慧养殖研究中，基于深度学习的猪只图像实例分割方法，是猪只个体识别、体重估测、行为识别等下游任务的关键。为解决模型训练需要大量的逐像素标注图像，以及大量的人力和时间成本的问题，采用弱监督猪只分割策略，制作弱监督数据集，提出一种新的特征提取骨干网络RdsiNet：首先在ResNet-50残差模块基础上引入第2代可变形卷积，扩大网络感受野；其次，使用空间注意力机制，强化网络对重要特征的权重值；最后引入involution算子，借助其空间特异性和通道共享性，实现加强深层空间信息、将特征映射同语义信息连接的功能。通过消融实验和对比实验证明了RdsiNet对于弱监督数据集的有效性，实验结果表明其在Mask R-CNN模型下分割的mAPSemg达到88.6%，高于ResNet-50、GCNet等一系列骨干网络；在BoxInst模型下mAPSemg达到95.2%，同样高于ResNet-50骨干网络的76.7%。而在分割图像对比中，使用RdsiNet骨干网络的分割模型同样具有更好的分割效果：在图像中猪只堆叠情况下，能更好地分辨猪只个体；使用BoxInst训练的模型，测试图像中掩码具有更高的精细度，这更有利于开展下游分析。 |
| WANG Haiyan JIANG Yehao LI Xuan MA Yunlong LIU Xiaolei |
| Shenzhen Institute of Nutrition and Health, Huazhong Agricultural University;Huazhong Agricultural University；Shenzhen Branch of Guangdong Laboratory of Lingnan Modern Agricultural Science and Technology;Shenzhen Institute of Nutrition and Health, Huazhong Agricultural University；Key Laboratory of Smart Farming for Agricultural Animals, Ministry of Agriculture and Rural Affairs;Key Laboratory of Agricultural Animal Genetics, Breeding and Reproduction, Ministry of Education |
| Key Words:pig weakly supervised instance segmentation spatial attention mechanisms involution operator |
| Abstract:In smart livestock farming research, deep learning based method for pig image instance segmentation is crucial for downstream tasks such as individual pig recognition, weight estimation, and behavior recognition. However, the model often requires a large number of pixel-wise annotated images for training, which imposes significant manpower and time costs. To address this issue, a weakly supervised pig segmentation strategy was proposed, creating a weakly supervised dataset, and introducing afeature extraction backbone network called RdsiNet. Firstly, the second-generation deformable convolution was incorporated into the ResNet-50 residual module to expand the network's receptive field. Secondly, spatial attention mechanisms were used to strengthen the network's weight values for important features. Finally, the involution operator was introduced to enhance deep spatial information and connect feature maps with semantic information by using its spatial specificity and channel sharing mechanism. The efficacy of RdsiNet for weakly supervised datasets was demonstrated through ablation experiments and comparative experiments. The experiments showed that the mean value of mask AP under the Mask R-CNN reached 88.6%, which was higher than a series of backbone networks such as ResNet-50 and GCNet.Meanwhile,the mean value of mask AP under the BoxInst reached 95.2%, which was also higher than that of ResNet-50 which reached only 76.7%. Furthermore, the display of image segmentation results of the test set showd RdsiNet also had better segmentation effect than ResNet-50. In the case of pig stacking, RdsiNet can better distinguish each pig. When using the BoxInst for training, RdsiNet can perfectly segment the outline of pigs, which was more conducive to downstream analysis. |
Transactions of the Chinese Society for Agriculture Machinery (CSAM), in charged of China Association for Science and Technology (CAST), sponsored by CSAM and Chinese Academy of Agricultural Mechanization Science(CAAMS), started publication in 1957. It is the earliest interdisciplinary journal in Chinese which combines agricultural and engineering. It always closely grasps the development direction of agriculture engineering disciplines and the published papers represent the highest academic level of agriculture engineering in China. Currently, nearly 8,000 papers have been already published. There are around 3,000 papers contributed to the journal each year, but only around 600 of them will be accepted. Transactions of CSAM focuses on a wide range of agricultural machinery, irrigation, electronics, robotics, agro-products engineering, biological energy, agricultural structures and environment and more. Subjects in Transactions of the CSAM have been embodied by many internationally well-known index systems, such as: EI Compendex, CA, CSA, etc.
|
| 下载PDF阅读器 |
|
|
|
