Research On Image Segmentation Based On Deep Neural Networks And Application In Assisting The Visually Impaired

With the rapid development of computer hardware lately,computer vision based on deep neural network has made a great progress.As a basic task of computer vision,semantic segmentation can realize scene perception.There are about 253 million visually impaired people in the world,encountering problems of life and travel indoors and outdoors.Based on the deep neural network,this paper designs the indoor and outdoor semantic segmentation network for the daily needs of visually impaired peopleIn indoor environment,there are problems like uneven illumination,insufficient illumination,object occlusion and object ambiguity.The RGB semantic segmentation model would be easily distracted by the difficulties above.With the development of depth cameras,models for indoor semantic segmentation have gradually promoted by combining depth images with RGB images.However,in indoor environments,the amount of information contained in RGB images and depth images is strongly affected by light exposure level.Moreover,the detail information represented by the RGB image,such as the shading of the texture,does not match the distribution pattern of the geometric morphology of the object represented by the depth images.The ACNet proposed in this paper can dynamically extract RGB-D features to solve the problems above.In the outdoor environment,the visually impaired people need a full range of environmental awareness technology.The semantic segmentation network of 360°panoramic circular image is designed.This model uses the network adaptation technology,trains on the ordinary data set,tests on the panoramic image,and achieves better performance.In addition,the skip connections with attention mechanism used in the model also enhance the model's ability to process image details.