Research On Small Object Semantic Segmentation Algorithm Based On Deep Neural Networks

Image semantic segmentation is an essential branch of scene understanding,and it has comprehensive application value and research significance,which can be applied to many fields such as automatic driving system,intelligent medical diagnostic system,virtual reality,etc.Traditional methods of semantic segmentation need to design manual features to construct the conditional random field model.This kind of methods is only applicable to small-scale datasets,but not to the large-scale and complex datasets which have been disclosed in recent years.Convolutional Neural Networks(CNN)can provide classifiers that are more powerful than traditional classification methods,and can automatically learn deep features,which significantly improves the accuracy of image semantic segmentation.However,these semantic segmentation methods based on CNN still have some challenges,such as the difficulty in segmenting the small objects in the complex scenes.In this thesis,three different semantic segmentation algorithms are designed based on the deep convolutional neural networks,aiming to solve the segmentation challenges of small objects.At the same time,on the premise of ensuring that the overall segmentation accuracy is better relative to current superior methods,it is more sensitive to the segmentation of small objects existing in images.The main contributions are as follows:1.Propose a semantic segmentation algorithm of the small-object-sensitive dual-channel convolutional neural networks.This algorithm designs a new cross-entropy loss function for semantic segmentation tasks.This function is more sensitive to the segmentation of small objects as the loss layer.Meanwhile,in order to improve the segmentation performance of other types of regions except for small objects,this algorithm uses the appropriate model fusion method to fuse two different models,making the final result not only sensitive to small object segmentation,but also ensures the overall segmentation performance,which has a certain improvement compared with the mainstream semantic segmentation methods.2.Propose a semantic segmentation algorithm of the small-object-sensitive end-to-end differential network.Based on the existing semantic segmentation network models,the algorithm has also designed a new segmentation model called "differential network"This network focuses on solving small object segmentation and other segmentation boundary problems.The two networks are combined into an end-to-end network for joint learning,so their parameters can be learned at the same time to obtain better solutions.The final model can be more sensitive to small objects segmentation on the basis of improving the overall segmentation accuracy.3.Propose a semantic segmentation algorithm for small objects combined with object detection.This work does not directly use a single neural network to segment both small-sized and large-sized objects simultaneously.Instead,it first obtains the bounding boxes of all small objects in the image through an object detection network.Then it uses small object image blocks as the input to design and train a small-object segmentation network model for low-resolution images,which can obtain small-object pixel-level segmentation maps,and finally uses the local segmentation results to modify the complete segmentation results of images,the modified segmentation maps has a better segmentation effect on small objects.