A Researh On Convolutional Neural Network Based Efficient Semantic Segmentation Method

Semantic segmentation is a very popular research direction in computer vision.It has great potential in video surveiliance,scene analysis,human-computer interaction and behavior analysis.Recently,semantic segmentation methods always use fully convolution network to construct their model,and take advantage of manually annotated data to train the model to get semantic segmentation model.However,the current semantic segmentation models often lead to complex and long running time in order to achieve high segmentation accuracy,which is difficult to be used in many scenarios with high speed requirements.Therefore,this paper studies the efficient semantics segmentation method with both high segmentation accuracy and fast inference speed.Firstly,we research on FCN,U-Net and DeepLabv3+.We compare the performance of these three models on Cityscapes dataset,and analyse the feature fusion way of these three models.To solve the problem of feature redundancy in deep learning model,an improved U-Net semantics segmentation model is proposed.Compared with the original U-Net,the network has fewer features,fewer parameters and higher segmentation accuracy on Cityscapes datasets.Secondly,by analyzing the task characteristics of target detection model and image recognition model,an efficient semantics segmentation method based on target detection transfer learning is proposed.This method uses YOLOv3 target detection model to transfer learning,designs loss function for the imbalance of training samples,and uses multi-type convolution module to extract high-level features to obtain more effective features and faster speed.The experimental results show that this method can effectively improve the segmentation efficiency while ensuring high segmentation accuracy.On the open test set Cityscapes,this method achieves the running speed of 0.171 s per frame and the segmentation accuracy of 79.1% mean-Intersection-over-Union.The experimental results show that this model can achieve better segmentation results than the model of transfer learning from image recognition.Furthermore,in order to alleviate the poor generalization ability of the model caused by sample bias,this paper proposes style transfer method to transfer the target image to the training image and then segment it.The experimental results show that the proposed method achieves a 5% improvement in accuracy compared with the original image segmentation.Thirdly,from the point of view of extracting and utilizing features more effectively,this paper proposes a fast semantic segmentation method based on multi-level convolution module.This method combines three different convolution kernels at different levels and fuses their extracted features to achieve more effective features with fewer model parameters.The experimental results show that the method has fast running speed and high segmentation accuracy.On the open test set Cityscapes,this method achieves 0.052 s running speed per frame and 75.2% mean Intersection-over-Union segmentation accuracy.We also apply the model to human body segmentation,and propose a multi-dataset joint training method to alleviate the bias of training samples and improve the robustness of the model.The experimental results show that our human body segmentation model can be effectively applied to various scenarios.Fourthly,we implement the two semantics segmentation models proposed in this paper on the embedded platform Xavier,and test the application of semantic segmentation based face style transfer.The results show that the proposed semantic segmentation model based on multi-level convolution module can run in real time on embedded platform.The experimental results show that our model can segment 640 * 360 * 3 images and transfer facial style at the speed of 0.172 s per frame on Xavier platform.In summary,this paper focuses on the efficient semantics segmentation method based on convolution neural network,and proposes an efficient semantics segmentation model from the perspectives of transfer learning from object detection,loss function design to alleviate the imbalance of training samples,extraction of multi-type high-level features,utilization of multi-level convolution module and implementation of embedded platform.The experimental results show that the proposed method can effectively improve the segmentation efficiency while ensuring the segmentation accuracy,and has the ability of practical application.