Research On Wall Defect Detection Based On Active Visual Method

In recent years,due to the rising labor costs and the rapid development of robotics,more and more industries are in urgent need of robots.China is a fast-developing devel-oping country.The people's longing for a better life is the basic requirement of everyone.The people's demand for precision-furnished houses is also getting higher and higher.The country has formulated a roadmap for the popularization of related technologies such as future green buildings,fully-furnished buildings,prefabricated buildings,and the use of prefabricated buildings.In this context,the construction mode of building decoration gradually shifts to the direction of specialization,mechanization and automation.The task of wall defect detection comes from the process of the Interior Finishing Robot's construction of the wall.Before the Interior Finishing Robot paints the wall,it needs to check the status of the wall and find some problems on the wall,such as convex,concave and cracks.Biological research has proven that observing things with the human eye does not distribute the attention evenly into the entire field of vision,but scans the eyes in a certain order and moves from one area to another.Grasp the area of interest.However,the current mainstream target detection algorithms are not used in this way,but are trained through full images containing accurate target bounding box to handle images passively.On the one hand,accurate labeling of target frames consumes a lot of resources? on the other hand,in some target detection tasks,accurate target frames are not required.Based on the above background,this paper conducts research around active vision models.By analyzing the research status of target detection algorithms,a model based on active vision method was proposed for the actual task of wall defect detection,and the following results were obtained:(1)An active vision method based on reinforcement learning is proposed.By adding the prior and multi-layer recurrent neural network to the recurrent attention model,the recurrent attention model is improved,so that the network achieves higher accuracy and faster convergence,and the network model can be applied to the actual environment.(2)A wall defect detection method based on attention model is proposed.Using a pre-trained attention model and using a memory and classification module based on a recurrent neural network,the network can be used on a mobile platform and can complete defect detection tasks while consuming a small amount of computing resources.(3)A method of wall defect detection based on active vision model is proposed.By using active vision methods based on reinforcement learning,multi-layer recurrent neural networks,Gaussian priors,and convolutional neural networks,the network can quickly detect wall defects in the robot's operating environment and achieve better accuracy.