Research On Image Depth Perception Based On Convolutional Neural Network

In recent years,the rapid development of artificial intelligence,such as virtual reality,augmented reality,self-driving cars and service robots,has increasingly appealed to depth visual technologies.Traditional depth perception is mostly the way of active hardware acquisition,such as radar based on optical reflection or electromagnetic wave reflection principle.However,such hardware is generally expensive and there are many restrictions.In recent years,the rapid development of convolution neural network has provided a new breakthrough for the development of visual depth technology.The existing research results can be divided into two methods: supervised and unsupervised.In this paper,we have studied the existing supervised depth perception algorithm based on convolutional neural network,and found that the feature parsing layers of existing models still has room for improvement.Therefore,in order to improve the analytical ability of the upper sampling module,this paper increases the resolution of the feature analysis module by increasing the depth and thickness of the upper sampling module.Compared with the existing upper sampling method,the improved upper sampling module not only enhances the correlation of features in the neighborhood of the sub-feature map,but also integrates the feature information in different size accepted domains in the feature mapping process,so as to make the final depth prediction result closer to the true depth distribution.In addition,in order to make the model converge to a better local optimal solution,the multi-resolution loss supervision information is added to the model training stage,which makes each upper sampling process have the effect of fine-tuning the coarse-grained and fine-grained depth information,and improves the final depth prediction accuracy of the model.For the existing unsupervised image depth prediction algorithm,this paper introduces a small amount of sparse supervision information in the model learning stage,which further constrains the solution space of the model and improves the depth prediction accuracy of the final model.For the improvements proposed in this paper,we have designed relevant comparative experiments to verify its effectiveness.For the improvement of supervised depth perception algorithm,we have validated the effectiveness of the upper sampling module and the effectiveness of multi-resolution loss supervision through multiple sets of contrast experiments on the NYU data set,and have obtained 0.545 linear mean square error and the depth prediction accurate pixel ratio of 79.5% on the NYU data set.For the improvement based on theunsupervised depth perception algorithm,we have validated the effectiveness of sparse depth supervision information on improving the model prediction accuracy on the KITTI data set,and obtained 4.211 linear mean square error and 86.2%accurate depth pixel ratio.