Research On High Precision Detection Of Road Diseases Based On Deep Learning

In recent years,the total mileage of highways is on the rise.Efficient and highquality maintenance of roads has become an urgent problem for smart transportation.Road surface diseases exist in different specifications,mainly represented by cracks and pits.If it is not surveyed and repaired in time,the daily travel of the people will bring a lot of inconvenience.However,traditional road disease identification and detection solutions are often costly,time-consuming,and have low accuracy,and lack commercial value for long-term development.At present,deep learning technology has achieved remarkable achievements,especially in the face recognition and other fields,which can have more higher recognition accuracy with faster detection efficiency.This topic takes road diseases as the research object,and the specific research work is as follows:(1)In order to achieve rapid localization and extraction of pavement diseases,the traditional Canny detection algorithm has the problem of easy loss of the key feature information when removing the noise points.Median filtering is put forward,which is easier to retain edge features.Aiming at the problem that the manual threshold is relatively random and troublesome,an optimization algorithm based on OSTU is designed to solve the threshold adaptively for the image.However,the disadvantages of using the traditional Canny algorithm to detect the pavement damage image in real environment are more obvious.(2)Aiming at the problem which large sets of data is required for deep learning model training,this paper proposes an algorithm to increase data based on morphological image transformation to enhance the model’s perception ability of border position and other features.Considering that under complex working conditions,the collected images may appear blurry and dim,such as foggy and cloudy days,a data augmentation method based on Retinex image enhancement algorithm is designed to enhance the abstraction of the model and improve the cognitive ability.(3)Aiming at the issues of conventional image processing algorithms such as difficulty in identifying road diseases and poor anti-noise ability,a road disease detection algorithm based on Faster RCNN deep convolution model is designed and adjust the network parameters to build a detection model suitable for the extraction of pavement disease features.This method has a better improvement than the traditional image processing algorithm.The m AP value can reach 47.98%,and its Recall is62.07%,and the processing of each frame of pictures reaches 59 ms.(4)Aiming at the problem of difficult identification and extraction of pavement diseases under complex working conditions,Faster RCNN network model which is more perfect is proposed.In order to integrate more semantic feature information,multiscale and multi-feature fusion is added to the backbone structure module,leaving detailed information and expanding the local receptive area of the high level.Considering the problem that it is difficult to extract features for diseases with small areas and inconspicuous geometric shapes,a multi-size and multi-proportion anchor box generation scheme is designed to suppress the proportion of missed detections.Aiming at low resolution or edge diseases that are prone to misdetection,the quantization decimal point algorithm of ROI Align with the bilinear interpolation method is proposed,which makes the disease features more reasonably match the spatial information.Experiments show that the m AP value of this method is improved by nearly 8.02%.