Background subtraction using deep long short-term memory neural networks and attention mechanism

Detecting moving objects is one of the important and practical issues in the field of machine vision. There are many solutions in this field.Some of these solutions are based on deep learning and deep neural networks, which are mainly supervised and offline.This paper presents an online and unsupervised approach leveraging deep learning for separating background from foreground in video data. The background is extracted as a low-rank matrix L using a deep neural network, then subtracting L from the original image gives the sparse foreground matrix.In designing the above neural network, a longshort-term memory (LSTM) network based on the attention mechanism has been used. The model usesunsupervised learning and can pay more attention and focus to the desired parts of the image.In order to evaluate the proposed model, the LASIESTA database, which covers a large number of challenges in the field of background subtraction, is selected.The proposed solution is compared with some well-known methods using standard criteria such as recall, precision and F-measure that shows 8, 10, and 5 percent improvement, respectively. Furthermore,it is qualitatively compared with the existing methods and succeeded in obtaining more favorable results.