Human Segmentation Based On Deep Leatning And Its Application

Human segmentation which is the human body from the image contains the segmentation,and according to the feature of the image itself to extract the human image information,the feature of which contains two aspects: 1,the image expressed by the scene;2,the human body where the image position.In daily life,no matter the intelligent monitoring system,the human-computer interaction,the scene analysis of the street and so on are inseparable from the computer vision technology,the computer understanding human body movement and the behavior will become the future computer vision technology an important development direction.Human image segmentation,as the basis of understanding and analysis of human behavior,has high requirements on the accuracy and real-time performance of segmentation.At present,most of human image segmentation methods can only guarantee one of the indexes,which makes the human image segmentation in practical application Subject to great restrictions.Before the deep learning is widely applied in the field of human image segmentation,the traditional human image segmentation method generally directly uses the image pixel value to operate.The segmentation effect is directly dependent on the setting of the threshold of the segmentation zero-crossing point,so its calculation is simple and efficient,But its limitation is also significant.The main reason is that the artificially designed features do not represent the image information very well and the robustness of human image segmentation with complex backgrounds is poor,which further limits the method In actual application.With the deepening of deep learning in the field of human image segmentation,practice has proved that its effect is obviously better than other non-depth learning methods.Considering the complexity of human image segmentation itself,there is still much research potential in this field.In this paper,by training the human body segmentation model under a variety of network structures,we compares and evaluates the application of different network structures in the field of human image segmentation,and then proposes a human image segmentation network that integrates the target region.This deep learning network not only utilizes Convolution neural network powerful feature extraction capabilities,the final output of the convolution feature is directly up-sampled input image features and each pixel mapping,to achieve the type of pixel prediction,and by fusion of the target area characteristics,the maximum limit The background area to avoid the interference of the segmentation results,improve the model performance.In Baidu human body image segmentation database and Pascal VOC database,the human image segmentation algorithm proposed in this paper is compared with the traditional human image segmentation algorithm.Under the same experimental data and hardware conditions,the experimental results are analyzed and demonstrated.Compared with some human image segmentation methods,a good balance is obtained between the overlapping rate and the real-time performance,and the effect of human image segmentation is greatly improved,showing good practical value.