Research Of Multi-scene Image Semantic Segmentation Based On Fully Convolutional Neural Network

In recent years,Machine Learning has been the main research of Artificial Intelligence.With the rapid development of Machine Learning,Computer Vision are also developing at a high speed.CNN(Convolutional Neural Networks)have been to be very effective methods for image recognition,image segmentation,and image localization after deep neural networks have been proposed.Whether in transportation,medical area,aviation,criminal investigation or daily life,Computer Vision is in a very important position.Therefore,the use of deep neural networks for building effective computer vision models has a high practical significance in these applications.Traditional image segmentation methods usually require a lot of complicated manual means and have low versatility.At the same time,the traditional image segmentation method has certain limitations in its effectiveness.With the proposing and development of deep neural networks,image semantic segmentation methods have been proposed with the needs of different applications.Image semantic segmentation is the pixel-level segmentation of images,that is,the classification of each pixel in the image.Image semantic segmentation based on deep neural network is more efficient and has better segmentation effect than traditional image segmentation method.Therefore,how to achieve and optimize the deep neural network applied to image semantic segmentation is one of the hot research spots.After a large number of related technologies and researches on related work,this thesis proposes and optimizes an image semantic segmentation method based on FCN(Full Convolutional neural Network).FCN changes the fully connected layer in the traditional convolution network to the convolution layer and combines with the deconvolution layer.The model uses the multi-scene image data set with the labels for model training,and applies the training model to pixel-level segmentation of images containing different targets.In the end,the model visualizes the test results by writing test modules and colors the segmentation result of the test set image.The data set used for training in the experiment,in addition to the existing open source data set,adds a large number of similar homemade images with labels.The experimental process uses two training modes with different parameters to achieve faster and better convergence,and uses mini batch images during training to accommodate the training of large data sets.Finally,through the comparison between the segmentation results of test set and the Ground Truth image,it is proved that the FCN training model has a higher validity and Robustness for segmentation of some targets in different scene images.