Study On Object Detection And Segmentation Based On Group Switchable Normalization

The switchable normalization smooths the loss optimization in the neural network training phase of the object detection and segmentation tasks by integrating three single methods,namely batch normalization,instance normalization and layer normalization through learnable weights,and realizes the accelerated training of the model.However,the learnable weights are determined by Softmax function.Each computation needs to use all the three single methods mentioned above at the same time.The sparseness is poor,there is computational redundancy and exist the influence of secondary methods on the main normalization,which limits the further improvement of object detection and segmentation accuracy by normalization features.In order to reduce the computational redundancy of normalization and improve the tasks accuracy.This thesis proposes an improved weight learning method of forward propagation to improve the sparseness of weight values without additional computation overhead,and to improve the influence of main normalization method.In addition,this thesis uses the group normalization with strong image information robustness to replace the layer normalization and propose the group switchable normalization method to improve the accuracy of traditional object detection and segmentation tasks.Furthermore,with the help of unsupervised domain adaptive learning,this thesis introduces adversarial training into the traditional object detection and segmentation models,combined with the group switchable normalization,further constrain the statistical distribution of features,smooth the optimization of training loss,and improve the average precision of cross domain object detection and segmentation tasks.The main research contents and achievements of this thesis are as follows:(1)Propose a forward propagation weight learning form of Logarithmicmax to replace the Softmax function in the original switchable normalization,so that the weight values can be 0 or 1,making the weight values more reasonable and the output distribution sparser.Moreover,the forward propagation does not need additional derivative gradient,and the computation is simple.In the loss optimization phase,the gradient descent will not be disturbed.(2)Use the group normalization with better robustness to image information to replace the layer normalization,and propose the group switchable normalization by combining with the improved weight learning method.In the object detection task,the test average precision of Faster R-CNN based on ResNet-50 and group switchable normalization in the open source Microsoft COCO 2014 validation set is increased to 34.1% compared with 32.9% under the condition of switchable normalization.In the same dataset,the test average precision of the instance segmentation task of Mask RCNN based on ResNet-50 and FPN is increased to 37.5% compared with 36.4% under the condition of switchable normalization.(3)With the help of unsupervised domain adaptive method,add the gradient reversal layers to the traditional object detection model Faster R-CNN to introduce adversarial training,combined with group normalization to further constrain the statistical distribution of source domain and target domain features,and smooth the loss optimization in the training phase.Increasing the cross domain object detection average precision of open source dataset SIM 10 k to Cityscapes from 39.7% to 44.1%,and Cityscapes to Foggy Cityscapes from 27.9% to 31.1%.(4)Improve the cross domain semantic segmentation model DA-DeepLab by introducing group switchable normalization with good robustness in small batch size to the discriminator network to constrain the statistical distribution of data,improves the cross domain semantic segmentation average precision of open source dataset GTA 5 to Cityscapes from 40.8% to 43.7%.In conclusion,this thesis proposes and proves that group switchable normalization can improve the accuracy of traditional object detection and instance segmentation tasks,and can be competent for cross domain object detection and semantic segmentation.