Research And Application Of Computer Vision Technology In Bridge Appearance Disease Identification

With the continuous development and improvement of China’s infrastructure construction,the construction of road traffic facilities,which is closely related to people’s life,has been paid more and more attention.The traditional construction mode requires manual construction or operation for any link of production,and the detection of bridge appearance diseases is no exception.However,the use of computer vision algorithms to identify the appearance of the building can make the detection of the disease away from the dependence on humans,and even reach the recognition rate that exceeds the manual detection.When using existing computer vision solutions to recognize the appearance of roads and bridges,it is often difficult to accurately recognize the original simple recognition task due to the complex and changeable appearance of the disease.In view of this,in order to solve the problem of difficult detection caused by the irregular shape of the disease,in this paper,the object detection and semantic segmentation models in computer vision are shared by the feature extraction network,so that the fusion model can obtain two kinds of recognition results,and achieve the purpose of optimizing the detection of bridge appearance diseases.Based on the above analysis,this paper mainly makes the following research:(1)Based on the theoretical basis of the target detection network Faster R-CNN and the image semantic segmentation network U-net,this paper proposes a fusion model based on Faster R-CNN and U-net that is suitable for solving the diversity of disease appearance.In order to make U-net better adapt to VGG-16 as its contraction path,a buffer convolution layer is used to adjust the number of channels of vgg-16’s feature output to the number of channels required by its expanding path after the feature extraction of vgg-16.At the same time,the jump connection of U-net on the contraction path and the expansion path is preserved,which helps U-net to better restore the original image information on the expansion path.(2)In order to train an effective fusion model,step-by-step training is adopted in the whole network training.The original loss functions of Faster R-CNN and U-net are still used in the model implementation,and the pre-training weights are used to initialize the model during training.After the pre-training weights are initialized,the parameters are selectively updated to achieve the effect of fast training of the model.(3)In the preprocessing of the data set,due to the feedback of the test results,the data set of the segmentation network for training the fusion model has been enhanced.The experimental results show that the segmentation network trained by data enhancement can obtain more accurate segmentation information.Based on the above analysis and research,it is shown that the fusion model can not only compensate for the insufficient position information of disease segmentation through object detection,but also solve the problem of discrete and irregular disease recognition by image segmentation.In addition,because the feature extraction network is shared between the models,the actual recognition of the fusion model to obtain dual recognition results is less than the recognition time of the disease through the two models,and this advantage will become more obvious when the image size increases.In the experiment,it is also found that the conventional data enhancement of the segmented training data of the fusion model is beneficial to increase the data set capacity and enable the model to obtain more accurate segmentation results.