研究目的
To develop a deep learning-based method for pixel-level material segmentation in images, addressing the challenge of varying appearances of the same material and improving accuracy over patch-based methods.
研究成果
The proposed SDC network effectively improves pixel-level material segmentation by maintaining feature resolution through dilated convolution and reducing artifacts via feature fusion. It outperforms existing methods in accuracy on the MINC dataset, with further improvements possible through CRF optimization and transfer learning.
研究不足
The MINC dataset has imbalanced labels, with some materials having fewer samples, which may affect recognition accuracy. The method relies on labeled data and may not generalize well to unlabeled regions or different datasets. The gridding effect from dilated convolution is mitigated but not fully eliminated.
1：Experimental Design and Method Selection:
The study uses a Skipped Dilated Convolution (SDC) network, which integrates dilated convolution layers to maintain spatial resolution and fuses features from bottom and top layers to mitigate artifacts. A VGG-16 network is used for feature extraction, and dense CRF is applied for optimization.
2：Sample Selection and Data Sources:
The MINC dataset with 23 material categories and 11,764 images is used; 7,061 images have labels, and additional 4,703 images were labeled for this study. Data augmentation (cropping, mirroring, rotating) is applied to increase training data.
3：List of Experimental Equipment and Materials:
Not specified in the paper.
4：Experimental Procedures and Operational Workflow:
The VGG network is pre-trained on MINC-2500 for patch classification, then transferred to the SDC network. Training uses augmented images resized to 512x512 with specific learning parameters. Testing is done on labeled subsets.
5：Data Analysis Methods:
Performance is evaluated using pixel accuracy, mean accuracy, and mean IoU metrics on the MINC dataset.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
