Weighted Feature Pyramid Networks for Object Detection

Li, Xiaohan; Lai, Taotao; Wang, Shuaiyu; Chen, Quan; Yang, Changcai; Lin, Jiazao; Fu, Zheng

doi:10.1109/ispa-bdcloud-sustaincom-socialcom48970.2019.00217

Cited by 68 publications

(35 citation statements)

References 8 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…When using a fixed‐size convolution for feature extraction, it is not efficient to extract long‐distance features 32 particularly for metal artifacts. A layered approach such as feature pyramid networks 33 is presently used to expand the receptive field, or the accuracy of the network is increased by deepening or widening the network. However, as the number of hyper parameters increases, the difficulty of network design and the computational overhead increase.…”

Section: Methodsmentioning

confidence: 99%

A semi‐supervised learning method of latent features based on convolutional neural networks for CT metal artifact reduction

et al. 2022

View full text Add to dashboard Cite

Purpose: X-ray computed tomography (CT) has become a convenient and efficient clinical medical technique. However, in the presence of metal implants, CT images may be corrupted by metal artifacts. The metal artifact reduction (MAR) methods based on deep learning are mostly supervised methods trained with labeled synthetic-artifact CT images. However, this causes the neural network to be biased toward learning specific synthetic-artifact patterns and leads to a poor generalization for unlabeled real-artifact CT images. In this study, a semisupervised learning method of latent features based on convolutional neural networks (SLF-CNN) is developed to remove metal artifacts while ensuring a good generalization ability for real-artifact CT images. Methods: The proposed semi-supervised method extracts CT image features in alternate iterations of a synthetic-artifact learning stage and a real-artifact learning stage. In the synthetic-artifact learning stage, SLF-CNN is fed with paired synthetic-artifact CT images and is constrained using mean-squarederror (MSE) loss and perceptual loss in a supervised learning fashion. In the real-artifact learning stage, the network weight is updated by minimizing the error between the pseudo-ground truths and the predicted latent features. The feature level pseudo-ground truths are obtained by modeling latent features using the Gaussian process. The overall framework of SLF-CNN adopts an encoder-decoder structure. The encoder is composed of artifact information collection groups to map the input artifact-affected synthetic-artifact CT images and real-artifact CT images into latent features. The decoder is composed of stacked ResNeXt blocks and is responsible for decoding latent features with high-level semantic information to reconstruct artifact-free CT images. The performance of the proposed method is evaluated through contrast experiments and ablation experiments. Results: The contrast experimental results indicate that the artifact-free CT images obtained by SLF-CNN have good metrics values, which are close to or better than those of typical supervised MAR methods. The metal artifacts in artifact-affected CT images are eliminated and the tissue structure details are preserved using SLF-CNN. The ablation experiment shows that adding realartifact CT images greatly improves the generalization ability of the network. Conclusions: The proposed semi-supervised learning method of latent features for MAR effectively suppresses metal artifacts and improves the generalization ability of the network.

show abstract

Section: Methodsmentioning

confidence: 99%

A semi‐supervised learning method of latent features based on convolutional neural networks for CT metal artifact reduction

et al. 2022

View full text Add to dashboard Cite

show abstract

“…In the feature fusion section of the neck network, the Dental-YOLO approach uses a feature pyramid network (FPN) to extract feature maps with different scales. The FPN combines top-down path convolution networks and lateral connections to develop high-level semantic feature maps at all scales [29]. An FPN can enhance object detection speed with high detection accuracy.…”

Section: A Datasetsmentioning

confidence: 99%

Dental-YOLO: Alveolar Bone and Mandibular Canal Detection on Cone Beam Computed Tomography Images for Dental Implant Planning

et al. 2022

View full text Add to dashboard Cite

In planning a mandibular posterior dental implant, identifying the exact location of the alveolar bone (AB) and mandibular canal (MC) is essential to determine the height and width of the available bone. Cone beam computed tomography (CBCT) is a 3D imaging modality widely used for dental implant planning, which requires a lower radiation dose compared to medical CT and can provide cross-sectional image quality to visualize AB and MC. The radiologist carried out the AB and MC detection processes manually on each section of the CBCT image until the appropriate area was determined for bone measurement. This process is time consuming, and the measurement accuracy depends on the ability and experience of the radiologist. This study proposes an automatic and simultaneous detection system for AB and MC based on 2D grayscale CBCT images, that can simplify and expedite dental implant planning. We introduce Dental-YOLO, an efficient version of YOLOv4 specifically developed to detect AB and MC, with two-scale feature maps at low and high scales. The height and width of the available bone in the implant area were estimated by using the detected bounding box attributes. The AB and MC detection performances using Dental-YOLO reached a mean average precision of 99.46%. The two-way analysis of variance (ANOVA) test showed no difference in the bone height and width measurements produced by the proposed approach and manual measurement by radiologists. Our results suggest that the Dental-YOLO detection system could be helpful for dental implant surgery and presurgical treatment planning.INDEX TERMS Alveolar bone, CBCT, bone measurement, dental implant planning, mandibular canal, object detection, YOLO.

show abstract

“…where is 4, and are the width and height of the region of interest, respectively, and the whole calculation result is rounded down to obtain the k, which means that layer k in the FPN is used to generate the feature patch. The particular structure of an FPN helps identify the location of small objects in the image, which renders feasibility and convenience to FPN training [50]. U-Net was used to classify the same data set with similar preprocessing procedures as FPN.…”

Section: Generation Of Ground Truth Datamentioning

confidence: 99%

Mapping of Subtidal and Intertidal Seagrass Meadows via Application of the Feature Pyramid Network to Unmanned Aerial Vehicle Orthophotos

Chen

Sasaki

2021

Remote Sensing

View full text Add to dashboard Cite

Seagrass meadows are one of the blue carbon ecosystems that continue to decline worldwide. Frequent mapping is essential to monitor seagrass meadows for understanding change processes including seasonal variations and influences of meteorological and oceanic events such as typhoons and cyclones. Such mapping approaches may also enhance seagrass blue carbon strategy and management practices. Although unmanned aerial vehicle (UAV) aerial photography has been widely conducted for this purpose, there have been challenges in mapping accuracy, efficiency, and applicability to subtidal water meadows. In this study, a novel method was developed for mapping subtidal and intertidal seagrass meadows to overcome such challenges. Ground truth seagrass orthophotos in four seasons were created from the Futtsu tidal flat of Tokyo Bay, Japan, using vertical and oblique UAV photography. The feature pyramid network (FPN) was first applied for automated seagrass classification by adjusting the spatial resolution and normalization parameters and by considering the combinations of seasonal input data sets. The FPN classification results ensured high performance with the validation metrics of 0.957 overall accuracy (OA), 0.895 precision, 0.942 recall, 0.918 F1-score, and 0.848 IoU, which outperformed the conventional U-Net results. The FPN classification results highlighted seasonal variations in seagrass meadows, exhibiting an extension from winter to summer and demonstrating a decline from summer to autumn. Recovery of the meadows was also detected after the occurrence of Typhoon No. 19 in October 2019, a phenomenon which mainly happened before summer 2020.

show abstract

Weighted Feature Pyramid Networks for Object Detection

Cited by 68 publications

References 8 publications

A semi‐supervised learning method of latent features based on convolutional neural networks for CT metal artifact reduction

A semi‐supervised learning method of latent features based on convolutional neural networks for CT metal artifact reduction

Dental-YOLO: Alveolar Bone and Mandibular Canal Detection on Cone Beam Computed Tomography Images for Dental Implant Planning

Mapping of Subtidal and Intertidal Seagrass Meadows via Application of the Feature Pyramid Network to Unmanned Aerial Vehicle Orthophotos

Contact Info

Product

Resources

About