基于改进SegNet的鸡只检测算法
Chicken detection algorithm based on improved SegNet
吉训生 1孙贝贝 1夏圣奎2
作者信息
- 1. 江南大学物联网工程学院,江苏无锡 214122
- 2. 南通天成现代农业科技有限公司兽医部门,江苏南通 226000
- 折叠
摘要
为实现智能化检测出鸡场中死亡鸡只,提出一种基于改进语义分割模型AT-SegNet的鸡只检测算法.基于对称编码解码结构SegNet,利用空洞卷积在解码前聚合不同感受野的上下文信息,设计一种三尺度注意力级联融合模块,以并联方式嵌入编、解码器间,丰富解码器信息.利用多层深度可分离卷积替代标准卷积,提取深层次语义信息,减少计算量提高实时性.将鸡群图像分割结果交并比与阈值对比判别鸡只状态.实验结果表明,改进的AT-SegNet较原算法的检测精度提高了 25.17％,能够在复杂鸡群环境中准确、高效地发现死亡鸡只.
Abstract
To realize intelligent detection of dead chickens in chicken farms,a chicken detection algorithm based on improved se-mantic segmentation model AT-SegNet was proposed.Based on the symmetric encoding-decoding structure SegNet,the hole convolution was used to aggregate the context information of different receptive fields before decoding,and a three-scale attention cascade fusion module was designed,which was embedded between the encoder and the decoder in parallel to enrich the decoder information.Multi-layer depth separable convolution was used to replace standard convolution,deep semantic information was extracted,the amount of calculation was reduced and the real-time performance was improved.The segmentation results of the chicken flock images were compared with the threshold to determine the status of the chickens.Experimental results show that the detection accuracy of the improved AT-SegNet is improved by 25.17％compared with the original algorithm,and it can accu-rately and efficiently detect dead chickens in a complex chicken flock environment.
关键词
深度学习/鸡只检测/语义分割/编码解码结构/注意力机制/软池化/深度可分离卷积Key words
deep learning/chicken detection/semantic segmentation/encoder-decoder structure/attention mechanism/soft poo-ling/depthwise separable convolution引用本文复制引用
基金项目
江苏省重点研发-现代农业基金项目(BE-2018334)
出版年
2024
NETL
NSTL
万方数据