Continuous Learning Semantic Segmentation Method For Robot Vision Based On Replay Strateg

Deep neural network-based semantic segmentation models for robot vision perform well when performing tasks that have already been trained.However,the real world where robots live is non-stationary,and the target data samples are often not available all at once.When the robot visual semantic segmentation model is incrementally trained to perform new tasks,the catastrophic forgetting of old knowledge is exacerbated by interference between old and new task parameters coupled with the phenomenon of background shift.In addition,data in real-world often cannot be stored due to privacy,security and other factors,in which case the robot vision semantic segmentation model fails.In order to address issues such as data privacy,alleviate catastrophic forgetting of robot vision semantic segmentation models in the process of continual learning,and reduce the resource cost of incremental model update knowledge,a continual learning semantic segmentation method for robot vision based on replay strategy is proposed.The main research is as follows:(1)To address issues such as data privacy and catastrophic forgetting when robot vision continual learning semantic segmentation models are confronted with the real world.Without storing old data,using generating adversarial network generation and web crawling as data sources,and label evaluation modules are used to solve unsupervised problems,while background self-inpainting modules are used to solve background shift,so that the samples meet training requirements.In this way,problems such as data privacy can be avoided,and catastrophic forgetting can be alleviated by replaying some old data during the continual learning process of the model.Finally,to further alleviate catastrophic forgetting,gating variables were introduced into the model and the special case of gating variables when combined with a continual learning replay strategy was investigated.It is used as a regularisation tool to increase model sparsity,reduce interference between old and new parameters during continual learning,and improve continual semantic segmentation accuracy.(2)For the problem of high spatial complexity of continual learning replay strategy models for semantic segmentation of robot vision.Based on the theoretical basis of research content(1),the sparse activation pattern of model neurons caused by the introduction of gating variables and the resulting gating mechanism are investigated by further utilizing the sparsity effect brought by gating variables to reduce the space complexity of continual learning replay strategy for robotic visual semantic segmentation.Finally,study and prove that on the basis of the experiment of research content(1),by adjusting the network sparsity to further reducing the inter-parameter interference,it is possible to achieve similar continual learning semantic segmentation effects using fewer data samples for training,thereby reducing the resource cost for training the replay strategy model.(3)Based on the previous two research contents,the robot vision continual learning semantic segmentation model based on the replay strategy is integrated and deployed on the robot platform for simulation testing.By building a robot platform and model deployment,from two settings of full sample and low cost,and by setting scenes of different complexity in the real world environment,the continual semantic segmentation ability of the model for robot vision RGB images is verified.