基于帧间高级特征差分的跨场景视频前景分割算法

Cross-Scene Foreground Segmentation Algorithm Based on High-Level Feature Differencing Between Frames

当前基于深度学习的有监督前景分割方法得益于大量待分割场景的标注信息,其性能大幅超越传统的无监督方法.然而,获取高精度的像素级标注需要耗费大量的人力和时间成本,这严重限制了有监督算法在无标注场景的部署应用.为解决对场景监督信息依赖的问题,设计了一种与传统的帧间差分法相融合的跨场景深度学习架构,即帧间高级特征差分算法.该算法重点围绕时域变化等跨场景共性知识的迁移,在不依赖待分割场景监督信息的前提下实现高精度分割.面向五类不同模式的困难场景开展实验,本文算法的平均F值达到0.8719,超越了当前最高性能的有监督算法FgSegNet_v2(相同的跨场景条件下)和最佳的无监督算法SemanticBS.本文算法对QVGA视频(320×240)的处理速度达到35帧/s,具有较好的实时性.

Benefiting from large amounts of ground-truths of to-be-segmented scenarios,deep-learning based and supervised foreground segmentation algorithms generally outperform conventional unsupervised methods.However,pixelwise annotation is a tedious task,especially when it comes to the annotation of foreground moving objects.This seriously limits the deployment of a supervised algorithm in a wide range of scenes without ground-truths.To address the dependence on supervised information of to-be-segmented unseen scenes,we design an inter-frame high-level feature differencing algorithm with a deep learning architecture via integrating the traditional frame differencing method.The proposed algorithm leverages the transfer of cross-scene common knowledge,such as temporal changes,so as to achieve high performance for the scene in the absence of supervised information of to-be-segmented scenes.We evaluate our method on five challenging scenes with different patterns.The average F-Measure of our algorithm is 0.8719,which surpasses the current highest-performance(supervised)algorithm(FgSegNet_v2)under the cross-scene learning condition and the best unsupervised algorithm SemanticBS.Our method which can process a QVGA(320×240)video at 35 frames per second shows favorable real-time performance.