High Precision Detection Technology Of Infrared Wall Cracks Based On Deep Learning

Concrete has the advantages of good performance and low price,so it is widely used in various modern structures.However,due to the characteristics of concrete pouring construction,it is difficult to control the quality of the material.In real life,factors such as excessive load,temperature change,salt-alkali corrosion,and construction technology may cause concrete components to dry out and creep and damage.Once a serious collapse accident occurs,it will cause huge loss of life and property.In the daily maintenance of concrete buildings,crack is one of the common and serious defects,which is the main cause of other diseases(water seepage,dust).Regular evaluation and inspection of concrete buildings are very important to protect people’s personal and property safety.Traditional methods of defect detection have problems such as low detection efficiency and poor safety.At the same time,manual inspection methods often rely on the experience level and professionalism of the inspectors,and their objectivity is poor.Therefore,how to detect the surface defects of concrete buildings in an efficient and convenient way is very important.In this paper,based on the deep learning algorithm and infrared thermal imaging technology to achieve high-precision detection of building wall cracks.This paper makes a detailed comparative analysis of the current classic target detection algorithms based on deep neural network,and uses data sets to test the performance of three commonly used algorithms.The experimental results show that the Single Shot Multibox Detector(SSD)balances the detection accuracy and speed well,and has obvious advantages in accuracy and reasoning speed,which meets the requirements of real-time detection.In this paper,a novel algorithm F-LIWC-DN(Lightweight Infrared Wall Crack Detection Network Based on FPN)based on deep neural network is proposed.The optimization direction and innovation of the proposed neural network include:(1)Optimizing the width-height-ratio and distribution of the priorbox of SSD network based on k-means clustering algorithm.Because the reasonable setting of priorbox greatly affects the detection performance of the final model,it is necessary to cluster the scale and distribution of cracks in the image to get the scale information of real cracks,so as to set the width-height-ratio of priorbox reasonably and improve the crack detection accuracy of the model.(2)The lightweight network MobileNetV2 is used as the backbone of the algorithm framework,and the deep separable convolution structure with fewer parameters is adopted to reduce the computational complexity and significantly improve the speed of model detection;(3)The feature pyramid structure is added,and the multi-level feature map is used to detect the target,and the low-level features are fully used to obtain the rich details of the object.Due to the trade-off relationship between the detection speed and accuracy,the use of MobileNetV2 to improve the network prediction speed will inevitably affect the performance of the model.FPN network predicts the object position and category information by fusing multi-layer feature map.The introduction of this method can improve the detection performance of the model while taking into account the reasoning speed of the algorithm.The collected infrared wall crack images are cut and flipped to ensure the diversity of data,and the VOC format data set is made by using annotation script.Based on the algorithm model proposed in this paper,the infrared wall crack data set is trained and tested.The average accuracy,the Params and the FLOPs are used to evaluate the performance of the model.Finally,the algorithm framework is optimized and improved according to the existing problems.The experimental data show that the crack detection accuracy of the improved F-LIWC-DN algorithm is 97%,and the model size is about 5M,which is only a quarter of the original SSD network model size.The algorithm proposed in this paper not only takes into account the speed of crack detection,but also ensures the accuracy of detection.It has a wide application prospect in the field of concrete building defect detection.