Image Semantic Segmentation Method Based On Attention Mechanism

Computer vision has been widely used in practical scenes.With the in-depth research and development of deep learning,computer vision has gradually replaced human vision to complete specific tasks.As one of many computer vision tasks,image semantic segmentation has a mature application in human-computer interaction,automatic driving and medical imaging.In recent years,researches on image semantic segmentation have flourished and solved many difficult problems in image semantic segmentation task from various angles.However,there are still some problems such as fixed receptive field,inconsistent semantic prediction and poor image detail prediction effect.In order to improve the overall performance of image semantic segmentation and promote the consistency of semantic prediction,this paper implements feature extraction of specific target objects based on deep learning method,which improves the performance of semantic segmentation to a certain extent.An image semantic segmentation network based on feature enhanced position attention module is proposed to solve the problem of fixed network receptive field.The network combines feature integration module.The feature integration module designed can help the network to fuse receptive fields of different spatial scales,further extract multi-scale feature maps,and achieve the purpose of extracting more abundant target features while increasing the receptive fields.After the feature integration module is connected to the backbone network,it can obtain more feature information as feature reserve for subsequent operations.The consistency of semantic prediction plays a decisive role in segmentation performance.Through a large number of literature research and experimental verification,it is found that the attention mechanism plays an important role in helping the network to capture rich context information,which can improve the prediction consistency of the network.Therefore,combining the attention mechanism and feature integration module,we propose an image semantic segmentation network that feature-enhance position attention module.Through experiments and evaluation on open data sets,the proposed network greatly improves the performance of image semantic segmentation.To solve the problem of detail information extraction and boundary contour prediction,an encoder-decoder image semantic segmentation network based on feature-enhanced attention is proposed.It helps the network to capture rich context information and simultaneously fuses boundary contour information,which is integrated into the network through feature enhancement.The encoder module is connected with the attention module to enhance the extraction of deep information.The connection feature integration module of decoder module strengthens the extraction of shallow information.After deep and shallow feature information is captured,feature information of high and low levels is fused through operations such as convolution and up-sampling to help the network obtain richer context information.Experiments and evaluations on open data sets verify the effectiveness of the proposed network,which greatly improves the performance of image semantic segmentation.