Reaearch On Background Subtraction Algorithm Based On Deep Learning Framework

With the continuous development of computer vision technology,video images have become an important medium for transmitting information.The demand for how to efficiently extract targets from massive video data is increasing.Background subtraction,as a key technology to achieve target detection,has become one of the important research topics in the field of computer vision.In the past three decades,domestic and foreign scholars have conducted in-depth research on background subtraction based on background modeling theories,and tried to extend it from experiments to practical applications.However,in actual applications,video scenes are complex and changeable,and it is difficult for traditional background subtraction algorithms to accurately extract foreground targets.In recent years,deep learning has been widely used in computer vision tasks due to its powerful representation capabilities for image data.Some researchers have tried to apply deep learning to background subtraction tasks to improve the generalization ability of background subtraction algorithms.Based on the theory of deep learning,this paper designed two background subtraction models that can effectively handle a variety of complex scenes:This paper proposes a multi-scale background subtraction model based on attention mechanism.In the field of image segmentation,the multi-scale structure can prompt the model to extract features of different scales and generate fine segmentation results.Therefore,some researchers try to use a multi-scale structure to construct a neural network model for background subtraction.However,the current background subtraction model based on multi-scale structure does not select different scale features when fusing them with different scale features.In responsing to this problem,this paper proposes to introduce the attention mechanism into the multi-scale structure,so that the model can produce multi-scale features of attention perception.In addition,this paper also designs multi-input encoder and multi-label supervised decoder,so that the model can extract more rich multi-scale features under the training mode of deep supervision.Evaluating the model performance with two public datasets,CDnet 2014 and LASIESTA,shows that the proposed model outperforms the state-of-the-art methods.In addition,this paper also proposes a background subtraction model based on conditional adversarial networks(CGAN,conditional Generative Adversarial Networks).With the continuous optimization of Generative Adversarial Networks(GAN)by researchers,GAN is more and more widely used in image segmentation tasks.Some researchers have applied it to background subtraction tasks.However,the current GAN-based subtraction models use the background image as a priori information,so the subtraction effect will be limited by the quality of the background image,and temporal information of the video sequence is ignored.To address the above problems,this paper proposes a CGAN-based background subtraction model,which is different from the previous GAN-based model.Two CGAN modules constitute static detection module and dynamic detection module respectively,static detection module takes a single frame image as a prior condition to extract the spatial features of the picture to generate static detection results.Dynamic detection module combines static detection results and adjacent video frames as a prior condition to aggregate the dynamic information of adjacent frames to generate the dynamic detection results.Experimental results on CDnet-2014 datasets demonstrate the high robustness of the proposed model.