Research And Application Of Image Segmentation By Deep Learning

Image segmentation technology is the most basic part in computer vision and is the basis of all other image processing methods.The quality of image segmentation technology will greatly affect the subsequent processing results.In the age of artificial intelligence,computer vision technology has been widely used in areas such as unmanned driving and security monitoring.The birth of these technologies is based on image segmentation.The research on image segmentation technology has been a decades-old history.From the early graphics processing algorithms to today’s full convolutional network algorithms,thanks to the development of hardware,the improvement of computing power,and the generation of mass image data in the information society.As well as the development of deep learning algorithms,computer vision technology represented by image segmentation has once again entered the field of vision.The image segmentation algorithm based on deep learning has been continuously proposed.Compared with the traditional image segmentation algorithms,there has been a qualitative leap in performance and effect,but there is still room for improvement.In this thesis,we improve on the basis of the classic U-net of the full convolutional neural network,combine the U-net with the hole convolution,and propose a new network model for image segmentation-the Unet combined with the hole convolution(Dilated Convolution Unet,DC-Unet).The network model that was originally applied to medical image segmentation was extended to more scene segmentation tasks.The network uses a convolutional encoder structure.The core of the network is mainly divided into two parts: the encoder and the decoder.The encoder encodes the input image and extracts features.The decoder decodes the extracted features and restores the semantics of the image.Before the output of the encoder is input to the decoder,a number of hole convolution operations are used to extract the features of the network coding at different scales and then fused,and a more expressive feature description is obtained,which is beneficial to the image Detection of tiny targets.The decoder and encoder structure are symmetric.In the decoder,the output of the previous network layer is combined with the network layer output of the corresponding position of the encoder as the input of the next network layer,and the shallow features extracted in the encoding stage are fully utilized,and the accuracy is more accurate.Positioning features,output more accurate segmentation results.In order to verify the actual segmentation effect of DC-Unet network model,this thesis tests and evaluates the two published image segmentation data sets and a self-made data set,and compares it with other classical image segmentation algorithms.The experimental results prove that the network has a good segmentation effect.