基于深度学习的监控视频树叶遮挡检测

Deep learning based approach for detecting leaf occlusion in surveillance videos

结合稀疏自编码器的自动提取数据特征能力和深度置信网络较好的分类性能,提出一种基于深度学习的监控视频树叶遮挡检测方法。首先从视频中随机选取一帧图像,通过栈式稀疏自编码器主动学习视频图像的特征信息,然后采用深度置信网络建立分类检测模型,最后引入学习速率自适应调整策略对整个神经网络进行微调。该方法不需要对视频连续取帧,具有较好的图像特征主动学习能力,克服了人工提取特征能力有限的缺陷。实验结果表明,在样本量充足的条件下,使用本文方法进行监控视频树叶遮挡检测可以达到88.97%的准确率。

Integrating the advantage of automatic feature extraction by sparse auto-encoder and the good classification performance of deep belief network, this paper proposes a detection approach for leaf occlusion in surveillance videos based on deep learning. Firstly, a frame is randomly selected from the video sequences, and a stacked sparse auto-encoder is used to actively learn the feature information in the video image. Next, a deep belief network is adopted to build a classification detection model. Finally, an adaptive learning rate strategy is introduced to fine-tune the whole artificial neural net- work. This method does not require consecutive video fetching frames and has better ability of active learning about image features, and therefore it overcomes the limitation of manual feature extraction. Experimental results demonstrate that the detection accuracy of the proposed method for leaf occlu- sion in surveillance videos can reach 88.97% under the condition of sufficient samples.