【目的】针对红外相机拍摄的野生动物图像数据量大、无效图像占比多、图像背景复杂等问题,提出一种可对图像进行自动、高准确率识别的模型,为生物多样性研究和野生动物保护工作提供更高效的支持。【方法】收集整理近4年来北京园林绿化...【目的】针对红外相机拍摄的野生动物图像数据量大、无效图像占比多、图像背景复杂等问题,提出一种可对图像进行自动、高准确率识别的模型,为生物多样性研究和野生动物保护工作提供更高效的支持。【方法】收集整理近4年来北京园林绿化生态系统监测网络各站点红外相机拍摄的约5 TB图像数据,对其手工标注并进行数据增强后自建10类共4234张图像数据集。基于ConvNeXt卷积神经网络,结合北京地区野生动物图像数据集特点,设计BSGG-ConvNeXt模型,使用BlurPool、SENet、全局响应归一化层(GRN)、GCNet提升模型识别能力,并在自建数据集上探究训练策略对ConvNeXt网络识别准确率的影响,通过与其他经典模型比较,明确BSGG-ConvNeXt模型的优势。利用公开的红外野生动物Snapshot Serengeti(SS)数据集和Caltech Camera Traps(CCT)数据集,验证模型的泛化能力。【结果】以ConvNeXt的ConvNeXt-T网络尺寸模型为例,其在自建数据集中的准确率为74.13%,乘加累积操作数(MACs)为4.47×10^(9)。应用不同改进方案发现,使用BlurPool后准确率提升2.2%,MACs降至1.07×10^(9);使用SENet后准确率提升3.2%;使用GRN并删掉缩放层后准确率升至87.18%,参数数量增至27.88×10^(6);使用GCNet后在不增大计算量的情况下准确率升至75.44%,但参数数量增至28.25×10^(6)。将上述改进方案结合得到的BSGGConvNeXt应用于ConvNeXt-T模型获得BSGG-ConvNeXt-T模型,参数数量虽有少量增多,但MACs降为1.07×10^(9),模型准确率升至83.63%,高于原模型。使用预训练权重后的BSGG-ConvNeXt-T模型准确率可达94.07%,高于ResNet-50(76.39%)、ResNeXt-50(87.60%)、MobileViT(90.00%)、DenseNet(87.66%)、RegNet(69.90%)、ConvNeXtv2(91.93%)、SwinTransformer的(86.23%)和MobileOne(71.53%),将BSGG-ConvNeXt模型应用于4种不同网络尺寸的ConvNeXt模型后,在自建数据集中的表现均优于未改进模型。BSGG-ConvNeXt模型在SS数据集中的识别准确率达50.28%,在CCT数据集中的识别准确率达56.15%,均高于原模型的准确率。【结论】BSGG-ConvNeXt模型识别红外相机拍摄的野生动物图像准确率更高,在自建、公开的野生动物红外图像数据集上均有较好表现,且具有一定泛化能力。展开更多
Activity recognition is a challenging topic in the field of computer vision that has various applications,including surveillance systems,industrial automation,and human-computer interaction.Today,the demand for automa...Activity recognition is a challenging topic in the field of computer vision that has various applications,including surveillance systems,industrial automation,and human-computer interaction.Today,the demand for automation has greatly increased across industries worldwide.Real-time detection requires edge devices with limited computational time.This study proposes a novel hybrid deep learning system for human activity recognition(HAR),aiming to enhance the recognition accuracy and reduce the computational time.The proposed system combines a pretrained image classification model with a sequence analysis model.First,the dataset was divided into a training set(70%),validation set(10%),and test set(20%).Second,all the videos were converted into frames and deep-based features were extracted from each frame using convolutional neural networks(CNNs)with a vision transformer.Following that,bidirectional long short-term memory(BiLSTM)-and temporal convolutional network(TCN)-based models were trained using the training set,and their performances were evaluated using the validation set and test set.Four benchmark datasets(UCF11,UCF50,UCF101,and JHMDB)were used to evaluate the performance of the proposed HAR-based system.The experimental results showed that the combination of ConvNeXt and the TCN-based model achieved a recognition accuracy of 97.73%for UCF11,98.81%for UCF50,98.46%for UCF101,and 83.38%for JHMDB,respectively.This represents improvements in the recognition accuracy of 4%,2.67%,3.67%,and 7.08%for the UCF11,UCF50,UCF101,and JHMDB datasets,respectively,over existing models.Moreover,the proposed HAR-based system obtained superior recognition accuracy,shorter computational times,and minimal memory usage compared to the existing models.展开更多
文摘【目的】针对红外相机拍摄的野生动物图像数据量大、无效图像占比多、图像背景复杂等问题,提出一种可对图像进行自动、高准确率识别的模型,为生物多样性研究和野生动物保护工作提供更高效的支持。【方法】收集整理近4年来北京园林绿化生态系统监测网络各站点红外相机拍摄的约5 TB图像数据,对其手工标注并进行数据增强后自建10类共4234张图像数据集。基于ConvNeXt卷积神经网络,结合北京地区野生动物图像数据集特点,设计BSGG-ConvNeXt模型,使用BlurPool、SENet、全局响应归一化层(GRN)、GCNet提升模型识别能力,并在自建数据集上探究训练策略对ConvNeXt网络识别准确率的影响,通过与其他经典模型比较,明确BSGG-ConvNeXt模型的优势。利用公开的红外野生动物Snapshot Serengeti(SS)数据集和Caltech Camera Traps(CCT)数据集,验证模型的泛化能力。【结果】以ConvNeXt的ConvNeXt-T网络尺寸模型为例,其在自建数据集中的准确率为74.13%,乘加累积操作数(MACs)为4.47×10^(9)。应用不同改进方案发现,使用BlurPool后准确率提升2.2%,MACs降至1.07×10^(9);使用SENet后准确率提升3.2%;使用GRN并删掉缩放层后准确率升至87.18%,参数数量增至27.88×10^(6);使用GCNet后在不增大计算量的情况下准确率升至75.44%,但参数数量增至28.25×10^(6)。将上述改进方案结合得到的BSGGConvNeXt应用于ConvNeXt-T模型获得BSGG-ConvNeXt-T模型,参数数量虽有少量增多,但MACs降为1.07×10^(9),模型准确率升至83.63%,高于原模型。使用预训练权重后的BSGG-ConvNeXt-T模型准确率可达94.07%,高于ResNet-50(76.39%)、ResNeXt-50(87.60%)、MobileViT(90.00%)、DenseNet(87.66%)、RegNet(69.90%)、ConvNeXtv2(91.93%)、SwinTransformer的(86.23%)和MobileOne(71.53%),将BSGG-ConvNeXt模型应用于4种不同网络尺寸的ConvNeXt模型后,在自建数据集中的表现均优于未改进模型。BSGG-ConvNeXt模型在SS数据集中的识别准确率达50.28%,在CCT数据集中的识别准确率达56.15%,均高于原模型的准确率。【结论】BSGG-ConvNeXt模型识别红外相机拍摄的野生动物图像准确率更高,在自建、公开的野生动物红外图像数据集上均有较好表现,且具有一定泛化能力。
基金funded by the Ongoing Research Funding Program(ORF-2025-890),King Saud University,Riyadh,Saudi Arabia.
文摘Activity recognition is a challenging topic in the field of computer vision that has various applications,including surveillance systems,industrial automation,and human-computer interaction.Today,the demand for automation has greatly increased across industries worldwide.Real-time detection requires edge devices with limited computational time.This study proposes a novel hybrid deep learning system for human activity recognition(HAR),aiming to enhance the recognition accuracy and reduce the computational time.The proposed system combines a pretrained image classification model with a sequence analysis model.First,the dataset was divided into a training set(70%),validation set(10%),and test set(20%).Second,all the videos were converted into frames and deep-based features were extracted from each frame using convolutional neural networks(CNNs)with a vision transformer.Following that,bidirectional long short-term memory(BiLSTM)-and temporal convolutional network(TCN)-based models were trained using the training set,and their performances were evaluated using the validation set and test set.Four benchmark datasets(UCF11,UCF50,UCF101,and JHMDB)were used to evaluate the performance of the proposed HAR-based system.The experimental results showed that the combination of ConvNeXt and the TCN-based model achieved a recognition accuracy of 97.73%for UCF11,98.81%for UCF50,98.46%for UCF101,and 83.38%for JHMDB,respectively.This represents improvements in the recognition accuracy of 4%,2.67%,3.67%,and 7.08%for the UCF11,UCF50,UCF101,and JHMDB datasets,respectively,over existing models.Moreover,the proposed HAR-based system obtained superior recognition accuracy,shorter computational times,and minimal memory usage compared to the existing models.
