| 基于显著性目标检测的图像前景提取算法研究 |
投稿时间：2024-04-15 修订日期：2024-06-06  点此下载全文 |
| 引用本文： |
| 摘要点击次数: 44 |
| 全文下载次数: 0 |
|
| 基金项目:国家自然科学基金项目（面上项目，重点项目，重大项目） |
|
| 中文摘要:为解决自然场景前景图像提取的背景复杂度高且依赖于人工交互的问题,提出通过检测显著性目标的方式实现提取前景图像的算法。该算法由显著性目标检测和Alpha估计两个子网络组成,前者在基准网络的基础上引入混合损失函数和深监督策略提高图像前景提取的完整度,后者通过增加上下文注意力引导模块的方式优化基准网络,实现对前景图像细节的精确恢复。将该算法改进思路用于扣像问题,算法在自制数据集上训练,公开数据集Composition-1K上测试,该算法在绝对误差和(MSE)指标上前景提取误差低至2.329×10-4平方像素,梯度误差(Grad)指标上误差值相较基准网络减小近60%。实验表明,该改进算法提高了提升网纱、玻璃等半透明区域的Alpha值估计的准确度从而提高前景图像提取精度,并且具有较强的鲁棒性与泛化能力,可应用于文本检测等下游任务。 |
| 中文关键词:图像前景提取 显著性目标检测 Alpha估计 混合损失函数 深监督 上下文引导 |
| |
| Research on Foreground Image Extraction Algorithm Based on Saliency Object Detection |
|
|
| Abstract:To address the challenges of high background complexity and reliance on manual interaction in natural scene image matting, an algorithm was proposed that automatically extracted foreground images by detecting salient objects. The algorithm consisted of two sub-networks: salient object detection and Alpha estimation. The former improved the completeness of foreground extraction by introducing a hybrid loss function and deep supervision strategy based on the baseline network. The latter optimized the baseline network by adding a context attention guidance module to accurately recover image details. The algorithm was trained on a self-made dataset and tested on the public dataset Composition-1K. Experimental results indicated that the algorithm achieved a matte error as low as 2.329×10-4 on the Mean Square Error (MSE) metric and reduced the error value on the Gradient (Grad) metric by nearly 60% compared to the baseline network. The experiment shows that the improved algorithm improves the accuracy of alpha value estimation in semi transparent areas such as mesh and glass, thereby improving the accuracy of image removal, and has strong robustness and generalization ability, , and can be applied to downstream tasks such as text detection. |
| keywords:Foreground image extraction Saliency object detection Alpha estimates Mixed loss function Deep supervision Context guided |
| 查看全文 查看/发表评论 下载pdf阅读器 |
位访问者,今日访问
次