Title page

会议/期刊：暂无

年份：2022

github链接：https://github.com/GewelsJI/DGNet

pdf链接： Public: https://arxiv.org/pdf/2205.12853.pdf Private暂无

Summary

We propose to excavate the texture information via learning the object level gradient rather than using boundary-supervised or uncertainty-aware modeling in Camouflaged object detection field.
- 基于边界监督或基于不确定性的模型，其通常响应的是伪装对象的稀疏边缘，从而引入噪声，特别是对于复杂的场景
- 很多伪装物的边缘无法定义或模糊不清
- 尽管伪装度很高，但仍有一些遗留下来的线索——we are interested in how the network mines these ‘discriminative patterns’ inside the object.
We propose a gradient-induced transition to automatically group features from the context and texture branches according to the soft grouping strategy.
DGNet outperforms existing state-of-the-art COD models by a large margin.
DGNet-S, runs in real-time (80 fps) and achieves comparable results to the cutting-edge model JCSOD-CVPR21 with only 6.82% parameters.
Application results also show that the proposed DGNet performs well in polyp segmentation, defect detection, and transparent object segmentation tasks.

Workflow

Methods

OverView

As discussed in Feature pyramid networks for object detection, 低维特征与高维特征一样重要
As suggested by Modnet: Real-time trimap-free portrait matting via objective decomposition, 不建议同时编码低维特征与高维特征.
综上，作者提出在伪装物识别中引入两个独立的编码器（内容编码器 context encoder和纹理编码器 texture encoder）

Context Encoder

使用了EfficientNet，提取5层金字塔特征
只取后三层富语义特征（affluent semantics
使用了两层堆叠 ConvBR $C_i$ × 3 × 3 → 降维，降低计算负担
The final outputs are three context features ${X^R_i}^5_{i=3} ∈ R^{C_i×H_i×W_i}$

Texture Encoder

作者构建了一个object-level gradient map监督的纹理编码器分支，用于补偿Context Encoder中后三层特征中的弱几何纹理特征

Object Gradient Generation

图像梯度：the directional change in an image’s intensity or color between adjacent positions.

考虑到全局图像梯度包含的背景噪声，作者构建了$Z_G$作为object-level gradient map，进行Texture Encoder的监督

Texture Encoder结构：Layer#04用于监督训练

Gradient-Induced Transition

根据通道数进行拆分组合，获得 $Q_i$

Soft-grouping策略

为了获得物体的多尺度信息，采用平行非线性投射到细粒度子空间的设计

Parallel Residual Learning

引入了平行残差思想

解码器

Neighbour connection encoder（NCD）

损失函数

$L_C$ 与 $L_G$ 均为加权IoU + BCE

训练策略

Adam + ImageNet预训练

The cosine annealing part of the SGDR strategy (I. Loshchilov and F. Hutter, “Sgdr: Stochastic gradient descent with warm restarts,” in ICLR, 2017) is used to adjust the learning rate, where the minimum/maximum learning rate and the maximum adjusted iteration are set to 10−5/10−4 and 20

Result-show

定量分析

定性比较

Efficiency Analysis

Ablation study

Downstream applications

启发和思考

深层特征与浅层特征先分离解码再分组、分尺度重整合
引入Gradient-induced mask，与传统的mask一起，多任务学习，有效提升了COD任务性能
如何对trade-off进行比较与撰写，很好的示范

Deep Gradient Learning for Efficient Camouflaged Object Detection

Paper-rading, pilyp