Learning High-quality Proposals for Acne Detection

Jianwei Zhang, Lei Zhang, Junyou Wang, Xin Wei, Jiaqi Li, Xian Jiang, Dan DuarXiv 2022 · OpenReview

TL;DR

This paper tackles the challenge of accurately detecting acne lesions, which are often small and have fuzzy boundaries, by improving the proposal generation stage of two-stage object detectors. They introduce a novel Spatial Aware Double Head (SADH) structure and a Normalized Wasserstein Distance (NWD) prediction branch to enhance proposal quality, leading to better acne detection performance.

What Problem Does This Paper Address?

Detecting and classifying acne lesions is a crucial first step in dermatological diagnosis and treatment. However, existing automated methods, particularly those using two-stage detection pipelines (like Faster R-CNN), struggle with the inherent characteristics of acne. These lesions often have irregular shapes and can be quite small, leading to a high number of "poor-quality" proposals generated by the Region Proposal Network (RPN). These low-quality proposals, which might have low confidence scores or inaccurate bounding boxes, can negatively impact the performance of the subsequent detection and refinement stages. The paper identifies two key issues contributing to this: a "flat classification confidence gradient" where RPN struggles to differentiate between easy and hard-to-detect lesions, and an "unsmooth label assignment" that creates a weak correlation between a proposal's classification score and its actual overlap (IoU) with a ground truth lesion.

Methods

The core of the proposed solution lies in enhancing the RPN's head structure.

  • Spatial Aware Double Head (SADH): This novel head disentangles the learning of classification and localization. For classification, it uses a 1x1 convolution to encourage spatial independence between adjacent feature map elements, leading to a steeper classification confidence gradient. For localization, it employs a 3x3 convolution to maintain spatial correlation, which is beneficial for accurate bounding box regression. This approach aims to make the RPN more discerning, better suppressing proposals for easy samples and improving the detection of harder ones.
  • Normalized Wasserstein Distance (NWD) Prediction Branch: To address the weak correlation between classification scores and IoUs, the authors introduce a new branch that predicts the Normalized Wasserstein Distance (NWD). NWD is argued to be a more suitable metric than IoU for evaluating the similarity of bounding boxes, especially for small objects like acne lesions. This predicted NWD is then used to "rectify" the original classification scores, creating a more reliable measure of proposal quality.
  • Soft-style Binary Cross Entropy (SBCE) Loss: The NWD prediction branch is trained using a novel SBCE loss function. This loss is designed to handle "soft labels" (continuous values) rather than just binary classifications, allowing for more nuanced learning of localization confidence.
  • Masked Crop: To handle high-resolution images efficiently and mitigate issues with partial lesions at image edges, the authors propose a "masked crop" preprocessing technique. This method crops images into smaller sub-images while masking out any partial lesions at the borders, preventing them from introducing inaccurate training signals.

Datasets

  • AcneSCU Dataset: The authors introduce a new, high-quality dataset named AcneSCU. It comprises 276 high-resolution facial images with 31,777 instance segmentation annotations across 10 lesion categories. This dataset is designed to be more precise and fine-grained than existing ones.
  • ACNE04 Dataset: The proposed method is also evaluated on the public ACNE04 dataset, which contains 1457 images with 18,983 bounding box annotations for a single lesion category.

Key Insights and Ideas

  • The importance of proposal quality: The paper strongly emphasizes that the performance of two-stage detectors is heavily reliant on the quality of the initial proposals. Improving RPN's output is a direct path to better overall detection.
  • Disentangling spatial learning: Separating the spatial learning for classification and localization within the RPN head is a key innovation. This allows for specialized processing that better suits each task.
  • NWD as a superior metric for small objects: The choice of NWD over IoU for localization confidence is a significant insight, particularly relevant for detecting small, challenging lesions.
  • Soft labels for better regression: Using SBCE loss for NWD prediction allows for a more robust training process by leveraging continuous ground truth values.
  • Masked crop for practical high-resolution processing: The masked crop technique offers a practical solution for handling large images in a computationally efficient manner while maintaining data integrity.

Key Results

  • On the AcneSCU dataset, the proposed method (Mask R-CNN + SADH + NWD) achieved a significant improvement in Average Precision (AP) of 1.0% over the baseline with SADH alone, and 2.6% over the base Mask R-CNN.
  • The SADH structure alone improved AP by 1.6% compared to the baseline.
  • The masked crop preprocessing method itself improved AP by 2.4%.
  • On the ACNE04 dataset, the proposed method outperformed previous state-of-the-art methods, achieving an AP of 0.22.
  • Ablation studies confirmed the effectiveness of each component: SADH improved AP, SBCE loss was better than L1 loss for NWD training, and NWD was a more effective localization metric than IoU, GIOU, and DIOU.
  • The paper also demonstrated that the proposed SADH and NWD components generalize well when integrated with other detection frameworks like HTC and Faster R-CNN.

Limitations and Future Work

  • False Detection Reasons: The authors acknowledge several reasons for false detections, including apparent lack of texture on cheeks, multiple tiny lesions being detected as one, very small lesions being missed due to harsh IoU thresholds, arbitrary lesion boundaries, and noisy labels. These are identified as subjects for future work.
  • Generalization to other skin conditions: While the method shows promise for acne, its generalization to other skin conditions with different lesion characteristics is not explicitly discussed.
  • Post-processing: The paper notes that post-processing techniques for merging detection results from sub-images were not the primary focus, suggesting potential for further optimization in this area.

Summary updated

Abstract

Acne detection is crucial for interpretative diagnosis and precise treatment of skin disease. The arbitrary boundary and small size of acne lesions lead to a significant number of poor-quality proposals in two-stage detection. In this paper, we propose a novel head structure for Region Proposal Network to improve the proposals' quality in two ways. At first, a Spatial Aware Double Head(SADH) structure is proposed to disentangle the representation learning for classification and localization from two different spatial perspectives. The proposed SADH ensures a steeper classification confidence gradient and suppresses the proposals having low intersection-over-union(IoU) with the matched ground truth. Then, we propose a Normalized Wasserstein Distance prediction branch to improve the correlation between the proposals' classification scores and IoUs. In addition, to facilitate further research on acne detection, we construct a new dataset named AcneSCU, with high-resolution imageries, precise annotations, and fine-grained lesion categories. Extensive experiments are conducted on both AcneSCU and the public dataset ACNE04, and the results demonstrate the proposed method could improve the proposals' quality, consistently outperforming state-of-the-art approaches. Code and the collected dataset are available in https://github.com/pingguokiller/acnedetection.
BibTeX
@misc{zhang2022learning,
  author = {Jianwei Zhang and Lei Zhang and Junyou Wang and Xin Wei and Jiaqi Li and Xian Jiang and Dan Du},
  title = {{Learning High-quality Proposals for Acne Detection}},
  year = {2022},
  eprint = {2207.03674},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2207.03674},
}
Figure