Research On Image Object Segmentation Based On Deep Convolution Neural Network

Image segmentation is one of the basic problems in the field of image processing and computer vision.Traditional image segmentation algorithms usually rely on low-level features such as color and texture to model,which are not ideal for image segmentation in complex scenes.Although the semantic segmentation algorithms based on deep convolutional neural network can automatically learn the appropriate feature representation for the current problem without manual selection and adjustment of features,there is less restricted by the low-level features,and they can obtain better precision than traditional methods could in many segmentation tasks.However,the algorithms still have many shortcomings in some segmentation tasks.(1)In the object segmentation task of sequence image or multi-view image,the traditional collaborative segmentation algorithms are not robust to complex multi-image segmentation,and the existing deep learning algorithms are easy to cause object segmentation error and segmentation inconsistency when there is large ambiguity in foreground and background.So we propose a multi-image segmentation algorithm which based on depth feature and fusing segmentation prior.To make the model learn the detail features from the multi-view images of complex scenes comprehensively,the PSPNet-50 network model is improved by integrating the high-resolution details of the shallow layer network,which also is used to reduce the effects of spatial information loss on the segmentation edge details as the network deepens.Afterward,one or two prior segmentations of the object are gained by using the interactive segmentation approach.These small priori segmentation integrations are fused into the new model.The network is then re-trained to solve the ambiguity segmentation problem between the foreground and the background and the inconsistent segmentation problem among multi-image.Finally,by constructing a fully connected conditional random field,the recognition ability of the deep convolutional neural network and the accurate locating ability of the fully connected condition random field are coupled together.The object region is effectively located,and the object edge is clearly detected.We evaluate our method on multi-image from various public data sets.Experiments show that our algorithm can exactly segment the region of the object in re-training classes while effectively avoiding the ambiguous region segmentation for those un-training object classes.Results show that our algorithm attains satisfactory scores not only in complex scene image sets with similar foreground and background contexts but also in simple image sets with obvious differences between the foreground and background contexts.The average scores of PA and IOU of our method are more than 0.95.Our algorithm has strong robustness in various scenes,and can achieve consistent segmentation in multi view images by fusing a small amount of priori integration.(2)Current deep convolutional neural network semantic segmentation models have too many weight parameters,large storage resource consumption and high computational complexity,which is not conducive to the application of mobile phones and other mobile terminals.And redesigning a compact network faces training problems such as under-fitting.So we propose a person image segmentation algorithm based on compressed depth network model.Firstly,the PSPNet-50 is fine-turned by using the person image segmentation dataset to obtain the person image segmentation model with higher precision.Then the convolutional level pruning and corresponding structure optimization are performed on the model after fine-turn.The parameters of the model are reduced a lot after that.Finally,the two-stage global filter level pruning strategy is used to prune the filter of the model.Compared with the method of layer-by-layer pruning and training,the retraining time is saved,and the accuracy of the model is guaranteed to a certain extent.As a result of that,the model parameters are further reduced.In addition,when training the network model,by adding auxiliary losses in the middle part of the network,the supervision training of the network is better realized,and the segmentation precision of the model is improved.Experiments show that the parameters of the person image segmentation model that get from our algorithm are reduced by 1/7.5 compared with the original model,the calculation amount is reduced to 1/6.6,the segmentation speed is increased by 2.4 times,and the runtime consumption memory is reduced to 1/1.7,the occupied storage space is reduced to 1/7.5,and the segmentation accuracy on the test set reaches a high level of 93.2%.Comparing with some commonly used segmentation models,our compressed model has many advantages.At the same time the algorithm of us achieves high segmentation accuracy on the character image,realizing compression and acceleration of the network model,and reducing the computing power requirements of the device and consumes the storage resources,and is beneficial to the mobile terminal and other mobile applications.It is advantageous to the application of the model in mobile terminals such as cellphone.