Human labor efficiency has become unable to keep the pace with gradually annual citrus increasing production.Highly efficient and intelligent citrus picking and accurate yield estimation is the key to solve the proble...Human labor efficiency has become unable to keep the pace with gradually annual citrus increasing production.Highly efficient and intelligent citrus picking and accurate yield estimation is the key to solve the problem.Success heavily depends on detection accuracy,prediction speed,and easy model deployment.Traditional target detection methods often fail to achieve balanced results in all those aspects.An improved YOLOv8 network model with four significant features is proposed.First,a lightweight FasterNet network structure was introduced to the backbone network,which reduced the number of parameters and computations while maintaining high-precision detection.Second,a progressive feature pyramid network AFPN structure was added to the neck network.Third,a parallel multi-branch attention mechanism PMBA was added before the detection head to improve the sensing ability after the feature fusion network.Fourth,a Wise-IoU was introduced to replace the original CIoU loss function to make the whole training process converge faster.Based on this,this study proposes an improved version of the YOLOv8 model:the FAP-YOLOv8.This improved model achieved an average accuracy(mAP@0.5)of 97.2%on the citrus datasets,with an accuracy that was 4.7%higher than the original YOLOv8,which was 19.2%,7.4%,5.1%,4.9%,and 5.2%higher than the other models:Faster R-CNN,CenterNet,YOLOv5s,YOLOx-s,and YOLOv7,respectively.The number of parameters was reduced by 55.45%,the computation was reduced by 20%compared to the YOLOv8 benchmark,and the frame rate reached 46.51 fps to meet the detection requirements of lightweight networks.The experiments showed that the FAP-YOLOv8 models all outperformed the comparison models.Consequently,the proposed FAPYOLOv8 model can help solve the citrus detection problem in orchards,which can be better applied to edge devices and provides strong support for intelligent orchard management.展开更多
Remote sensing image(RSI)with concurrently high spatial,temporal,and spectral resolutions cannot be produced by a single sensor.Multisource RSI fusion is a convenient technique to realize high spatial resolution multi...Remote sensing image(RSI)with concurrently high spatial,temporal,and spectral resolutions cannot be produced by a single sensor.Multisource RSI fusion is a convenient technique to realize high spatial resolution multispectral(MS)images(spatial spectral fusion,i.e.SSF)and high temporal and spatial resolution MS images(spatiotemporal fusion,i.e.STF).Currently,deep learning-based fusion models can only implement SSF or STF,lacking models that perform both SSF and STF.Multiresolution generative adversarial networks with bidirectional adaptive-stage progressive guided fusion(BAPGF)for RSI are proposed to implement both SSF and STF,namely BPF-MGAN.A bidirectional adaptive-stage feature extraction architecture infine-scale-to-coarse-scale and coarse-scale-to-fine-scale modes is introduced.The designed BAPGF introduces a previous fusion result-oriented cross-stage-level dual-residual attention fusion strategy to enhance critical information and suppress superfluous information.Adaptive resolution U-shaped discriminators are implemented to feed multiresolution context into the generator.A generalized multitask loss function unlimited by no-reference images is developed to strengthen the model via constraints on the multiscale feature,structural,and content similarities.The BPF-MGAN model is validated on SSF datasets and STF datasets.Compared with the state-of-the-art SSF and STF models,results demonstrate the superior performance of the proposed BPF-MGAN model in both subjective and objective evaluations.展开更多
基金financially supported by the Yunnan Provincial Major Science and Technology Special Project:Research and Development and Application Demonstration of Key Technology for Digitization of Cloud Fruit(Grant No.202002AE09001002).
文摘Human labor efficiency has become unable to keep the pace with gradually annual citrus increasing production.Highly efficient and intelligent citrus picking and accurate yield estimation is the key to solve the problem.Success heavily depends on detection accuracy,prediction speed,and easy model deployment.Traditional target detection methods often fail to achieve balanced results in all those aspects.An improved YOLOv8 network model with four significant features is proposed.First,a lightweight FasterNet network structure was introduced to the backbone network,which reduced the number of parameters and computations while maintaining high-precision detection.Second,a progressive feature pyramid network AFPN structure was added to the neck network.Third,a parallel multi-branch attention mechanism PMBA was added before the detection head to improve the sensing ability after the feature fusion network.Fourth,a Wise-IoU was introduced to replace the original CIoU loss function to make the whole training process converge faster.Based on this,this study proposes an improved version of the YOLOv8 model:the FAP-YOLOv8.This improved model achieved an average accuracy(mAP@0.5)of 97.2%on the citrus datasets,with an accuracy that was 4.7%higher than the original YOLOv8,which was 19.2%,7.4%,5.1%,4.9%,and 5.2%higher than the other models:Faster R-CNN,CenterNet,YOLOv5s,YOLOx-s,and YOLOv7,respectively.The number of parameters was reduced by 55.45%,the computation was reduced by 20%compared to the YOLOv8 benchmark,and the frame rate reached 46.51 fps to meet the detection requirements of lightweight networks.The experiments showed that the FAP-YOLOv8 models all outperformed the comparison models.Consequently,the proposed FAPYOLOv8 model can help solve the citrus detection problem in orchards,which can be better applied to edge devices and provides strong support for intelligent orchard management.
基金funded by the National Key Research and Development Program of China under Grants 2020YFB2104400 and 2020YFB2104401the National Natural Science Foundation of China under Grant 82260362the Hainan Major Science and Technology Program of China under Grant ZDKJ202017.
文摘Remote sensing image(RSI)with concurrently high spatial,temporal,and spectral resolutions cannot be produced by a single sensor.Multisource RSI fusion is a convenient technique to realize high spatial resolution multispectral(MS)images(spatial spectral fusion,i.e.SSF)and high temporal and spatial resolution MS images(spatiotemporal fusion,i.e.STF).Currently,deep learning-based fusion models can only implement SSF or STF,lacking models that perform both SSF and STF.Multiresolution generative adversarial networks with bidirectional adaptive-stage progressive guided fusion(BAPGF)for RSI are proposed to implement both SSF and STF,namely BPF-MGAN.A bidirectional adaptive-stage feature extraction architecture infine-scale-to-coarse-scale and coarse-scale-to-fine-scale modes is introduced.The designed BAPGF introduces a previous fusion result-oriented cross-stage-level dual-residual attention fusion strategy to enhance critical information and suppress superfluous information.Adaptive resolution U-shaped discriminators are implemented to feed multiresolution context into the generator.A generalized multitask loss function unlimited by no-reference images is developed to strengthen the model via constraints on the multiscale feature,structural,and content similarities.The BPF-MGAN model is validated on SSF datasets and STF datasets.Compared with the state-of-the-art SSF and STF models,results demonstrate the superior performance of the proposed BPF-MGAN model in both subjective and objective evaluations.
