Source-linked AI summary
Dense Extreme Inception Network: Towards a Robust CNN Model for Edge Detection
Xavier Soria, Edgar Riba, Angel D. Sappa
TL;DR
Edge detection remains an open problem, especially when available datasets omit edges or mix edge annotations with object contours. The paper introduces DexiNed and BIPED, reporting cross-dataset improvements while noting that one evaluation dataset contains object contours in its ground truth.
Problem
Edge detection remains open, and existing datasets can contain missing edges or annotations intended for boundaries and contours rather than edges.
Method
The paper proposes DexiNed, a CNN with dense multi-scale feature processing and upsampling blocks that produces thin edge-maps, trained using BIPED.
Results
DexiNed, trained once on BIPED, outperforms state-of-the-art approaches when evaluated on other edge-oriented datasets.
Takeaways & Limitations
BIPED provides a carefully annotated edge-detection dataset, while DexiNed supports evaluation across other edge-oriented datasets without retraining on them.
Takeaways & Limitations
CID contains only 40 images, and its ground-truth maps include object contours in addition to edges.
Abstract
from arXiv · showhide
This paper proposes a Deep Learning based edge detector, which is inspired on both HED (Holistically-Nested Edge Detection) and Xception networks. The proposed approach generates thin edge-maps that are plausible for human eyes; it can be used in any edge detection task without previous training or fine tuning process. As a second contribution, a large dataset with carefully annotated edges has been generated. This dataset has been used for training the proposed approach as well the state-of-the-art algorithms for comparisons. Quantitative and qualitative evaluations have been performed on different benchmarks showing improvements with the proposed method when F-measure of ODS and OIS are considered.
1. Introduction
Edge detection remains an open problem despite its broad use and many existing methods, partly because commonly used datasets can omit edges. The paper introduces a dedicated edge dataset and a robust CNN detector.
- Edge detection supports classical and modern computer-vision processes, including segmentation, recognition, image-to-image translation, and photo sketching.
- Existing boundary-detection and object-segmentation datasets may contain missing edges, making training difficult and causing predicted edge-maps to lose edges.
- The paper targets cross-dataset generalization, evaluating a model on edge-detection datasets without training on those evaluation sets.
- BIPED is a released dataset with carefully annotated edges for training and evaluating edge-detection models.
- DexiNed is a robust CNN edge detector trained from scratch without pretrained weights.
2. Related Work
Prior edge detectors span low-level, biologically inspired, classical-learning, and deep-learning approaches. DexiNed is presented as a multi-scale CNN architecture related to the deep-learning line of work.
- Edge-detection methods are categorized into low-level feature, brain-biologically inspired, classical-learning, and deep-learning algorithms.
- Low-level feature: Low-level methods commonly smooth images with Gaussian filters or manually designed kernels, often ending with non-maximum suppression.
- Brain-biologically inspiration: Biologically inspired methods model visual edge or contour formation using mechanisms such as simple cells, Gabor filters, and derivative-of-Gaussian filters.
- Classical learning algorithms: Classical-learning approaches include sparse and dictionary learning, gPb, and structured forests, but remain limited in challenging scenarios.
- Deep learning algorithms: Deep-learning approaches use CNNs to predict edge-maps, with HED introducing multi-scale outputs and later methods proposing improvements.
- Proposed architecture: DexiNed uses six encoder blocks with intermediate upsampled edge-maps that are concatenated to produce a fused edge-map.
3. Dense Extreme Inception Network for Edge Detection
DexiNed combines a dense extreme inception network with upsampling blocks to produce same-resolution, thin edge-maps from RGB images. Multi-scale outputs are fused and deeply supervised with weighted cross-entropy.
- Architecture: DexiNed uses a six-block encoder inspired by Xception, plus upsampling blocks that generate intermediate edge-maps and a final fused prediction.The encoder receives RGB images, while each upsampling block receives feature maps from a Dexi block.
- Architecture: The architecture uses standard convolutions, adds a block for 2D edge-map output instead of Xception’s exit flow, and employs two types of skip connections.DexiNed also reduces filters in blocks 4 and 5 from 728 to 512.
- Architecture: Since block 3, edge-connections average sub-block outputs to preserve important edge features lost through deep convolutional processing.This design is motivated by the insufficiency of a single main connection for retaining edge information.
- Upsampling Block: The upsampling block uses conditional convolutional and deconvolutional sub-blocks to restore feature-map scale and produce thin edge-maps.The first sub-block handles scale differences of 2, while the second is iterated for larger differences.
- Outputs and Loss: Eight edge-map predictions are produced, concatenated for fusion, and supervised using weighted cross-entropy across scales.The final fusion output is one member of the predicted edge-map set, and the loss weights edge and non-edge ground-truth pixels.
4. Experimental Setup
The experiments introduce BIPED as a carefully annotated edge-detection benchmark and evaluate DexiNed across edge and boundary datasets. Performance is measured with ODS, OIS, AP, and F-measure, using a defined training protocol.
- BIPED: BIPED contains 250 outdoor 1280×720 images annotated and cross-checked by computer-vision experts, and is released as an edge-detection benchmark.Its construction addresses the scarcity and annotation errors of existing edge datasets.
- Test Datasets: The evaluation includes MDBD for edge detection and CID, BSDS, NYUD, and PASCAL for object contour, boundary, or segmentation-related evaluation.DexiNed is evaluated on selected test subsets, including 20 MDBD images, all CID images, NYUD’s testing set, and 505 PASCAL images.
- Metrics: ODS, OIS, and AP are used as evaluation metrics, with F-measure reported for fixed contour threshold and per-image best threshold settings.BIPED provides the annotated edge maps used as ground truth for these metrics.
- Implementation: Training uses TensorFlow, Adam, a 10^-4 learning rate, batch size 8, 400x400 color inputs, and convergence after 150k iterations.The reported training process takes around two days on a TITAN X GPU.
- Evaluation Design: Table 1 evaluates eight DexiNed predictions on BIPED and compares state-of-the-art methods trained and evaluated with BIPED.The table caption distinguishes within-model prediction evaluation from comparative model evaluation.
5. Experimental Results
DexiNed is evaluated through upsampling and edge-map fusion choices, comparisons on BIPED, and cross-dataset testing. The results support strong edge-detection performance while showing that boundary-oriented datasets can penalize edge predictions.
- Upsampling selection: DexiNed-dc is selected because its precision/recall curves show a small performance advantage, although all three upsampling versions have similar F-measures.DexiNed-dc uses trainable transpose-convolution kernels; DexiNed-bdc uses bilinear-initialized kernels, and DexiNed-sp uses subpixel convolution.
- Edge-map outputs: DexiNed-a achieves the best quantitative results across the three evaluation metrics, while DexiNed-f is qualitatively better despite slightly lower quantitative scores.Both fused and averaged edge-maps are retained as the main comparison outputs, with DexiNed later referring to DexiNed-f.
- BIPED comparison: DexiNed-a reaches the best results in all evaluation metrics on BIPED, while DexiNed-f and DexiNed-a are generally the top-performing approaches.For recall above 75%, DexiNed has the best F-measure; RCF and BDCN have similar quantitative results but qualitatively weaker outputs.
- Cross-dataset evaluation: DexiNed is trained once on BIPED and obtains the best performance on MDBD, but not on BSDS500, NYUD, or PASCAL.The latter datasets were not intended for edge detection, so their metrics penalize edges predicted by DexiNed.
- Annotation effects: On fully annotated images, F-measure exceeds 80% in BSDS500, BSDS300, and NYUD, whereas incomplete annotation can reduce it to about 30%.For MDBD, the worst F-measure exceeds 75%, highlighting differences between edge detection and contour or boundary detection.
6. Conclusions
The paper presents DexiNed as a deep structured model for image edge detection and reports strong generalization after training only on BIPED. It also releases a carefully annotated edge dataset and identifies contour and boundary detection as future work.
- Conclusions: DexiNed is proposed as a deep structured model that generates thin edge-maps for image edge detection.The paper describes it as the first DL-based approach able to generate thin edge-maps.
- Conclusions: After one training run on BIPED, DexiNed outperforms state-of-the-art approaches on other edge-oriented datasets.The conclusion reports this result as evidence supporting the proposed approach.
- Conclusions: BIPED is released as a carefully annotated dataset for edge detection.The dataset is shared with the research community as a separate contribution.
- Future work: Future work will address contour and boundary detection using the proposed architecture and approach.