Research And Implementation Of Target Segmentation Algorithm Based On Deep Learning

Semantic segmentation,as the foundation of computer vision,is a pixel-level task of predicting different semantic categories in images.It is widely used in human-computer interaction,autonomous driving,and scene understanding.Deep learning has greatly improved the performance of semantic segmentation,but general segmentation models have large storage costs and low speed due to convolution stacking and multi-scale feature aggregation.Based on the problems of large space occupation and unbalanced accuracy and speed in existing segmentation algorithms,semantic segmentation algorithms applied on an embedded platform are proposed.They contain a segmentation algorithm for pupil object and a segmentation algorithm for urban street scene object.The specific research and innovations are as follows:(1)First,the pupil semantic segmentation dataset is constructed.It contains the Eye Lab dataset and the Open EDS dataset.In the Eye Lab dataset,the images are acquired by the laboratory face acquisition equipment,and the labels are obtained by manually labeling the pupil object with Labelme software.The images in the Eye Lab dataset contain the eye region and other face information.They ensure that pupil object are well detected in distant scenes.In the public Open EDS dataset,the images only contain close-up eye regions and the label categories are divided into pupil,sclera,and iris.Therefore,the labels are modified and only the pupil object is retained.Eye Lab and Open EDS constitute a complete pupil segmentation dataset.(2)A small real-time segmentation algorithm called LRPNet is proposed based on pupil object.A bilateral network is used in the algorithm,it contains a spatial branch network and a semantic branch network.The spatial branch network with shallow layers and wide channels captures spatial details while generating high-resolution feature representations.The semantic branch network with deep layers and narrow channels is used to obtain contextual semantic information with a large receptive field.An effective feature fusion module is designed to combine two different pieces of information.At the same time,the images are enhanced by horizontal flipping,translation operations,Gaussian blurring,and sharpening operations during the training process.They improve the robustness of the model.The accuracy and data processing speed of the LRPNet algorithm are high on both Open EDS and Eye Lab datasets.To further reduce the number of parameters,the depthwise separable convolution is used to optimize it.So a more lightweight LRPNet-ds algorithm is obtained and the storage capacity of the model is reduced from 0.08 MB to 0.04 MB.(3)A higher-accuracy semantic segmentation algorithm called MEF-SCNN is proposed for urban street scene object based on the lightweight real-time algorithm Fast-SCNN.The spatial attention mechanism(SAM)is applied in the feature fusion module.It solves the problem of blurred boundaries in the segmentation area and makes the model learn more useful information about the spatial location.The channel attention mechanism(CAM)is used in the classifier module.It solves the problem that different targets with similar features are easily confused.And the ability to extract semantic features is improved in the network.The boundary guiding mechanism(BGM)is applied in the training phase to strengthen the learning of shallow detail features in the network.It further improves the accuracy.Compared with Fast-SCNN,the algorithm improves the accuracy by 4.12%.