Research On Image Semantic Segmentation Based On Full Convolution Neural Network

Image semantic segmentation is one of the hot research topics in computer vision,playing a significant role in recognizing and understanding the content of images.Semantic segmentation is a dense classification task that aims to classify each pixel in an image according to a set of semantic labels,dividing the image into several pixel regions with specific semantics.In recent years,with the improvement of computer software and hardware performance and the emergence of fully convolutional neural networks,the development of image semantic segmentation technology has entered a new stage.The accuracy of semantic segmentation models has been significantly improved,and they have been widely applied in fields such as autonomous driving and medical image analysis.However,existing image semantic segmentation methods based on fully convolutional neural networks still suffer from problems such as feature shift leading to poor fusion of different scale features,inadequate utilization of spatial and contextual information,and slow inference speed.To address these problems,this paper optimizes the semantic segmentation model from multiple aspects,with the main research content as follows:(1)To address the problem of poor multi-scale feature fusion caused by feature shift,this paper proposes a semantic method based on multi-scale feature alignment and aggregation.Currently,semantic segmentation models mostly use feature addition or feature concatenation to aggregate high-level semantic information and shallow spatial information.However,such methods ignore the misalignment between feature maps at different levels,which may lead to incorrect classification of small object boundaries.This paper strengthens the focus on feature fusion and proposes a multi-scale feature alignment and aggregation module to align and aggregate features at different levels,achieving more refined segmentation results.This method gradually fuses shallow features of different resolutions to restore spatial details,integrates feature information from low to high resolutions,and gradually achieves feature refinement.Experimental results show that the proposed method can effectively improve the problem of feature misalignment and improve the accuracy of image segmentation.(2)In response to the problem of insufficient utilization of spatial and contextual information and slow inference speed caused by large network models,an efficient semantic segmentation network,BCENet,is proposed based on boundary guidance and context information.To improve the inference speed,the network structure is optimized in this paper,using a lightweight feature extraction network to reduce the model size and improve the algorithm’s running efficiency.To refine the segmentation effect,this paper improves the model’s segmentation accuracy from two aspects: enhancing context information and spatial information.While ensuring segmentation efficiency,BCENet introduces cross-attention mechanism and multi-scale pyramid pooling module in the decoding process to capture the context correlation between pixels and obtain context information of different scales,improving the recognition ability of objects of different sizes.In terms of spatial information,a boundary assistance module is designed to supervise and guide the edge information of low-level features,providing more accurate detail information for image segmentation.Experimental results show that BCENet has good segmentation accuracy while ensuring real-time segmentation.(3)Autonomous driving is one of the most important application scenarios for image semantic segmentation.Aiming at the drivable area detection task in the autonomous driving scene,a road drivable area segmentation system is designed and implemented.This system applied the two semantic segmentation methods mentioned in this paper and built a visual display platform for the drivable area based on the segmentation algorithm.The system was tested in various road scenarios and at different times,and the results showed that the road drivable area segmentation system designed in this paper was feasible and effective,and had certain application value.