Research On High Ratio Image Compression Algorithms For Tunnel Vision Inspection

The traditional tunnel maintenance method mainly relies on experienced workers carrying a ladder truck to complete.This kind of method is not only slow and inefficient,but also has a high risk factor,which cannot effectively guarantee the personal safety of the workers.With the development of intelligent manufacturing technology,vision-based tunnel maintenance methods have gradually become a new trend,and gradually replaced traditional manual maintenance method.However,it is noticeable that there are enormous images available from visual tunnel inspections,which inevitably leads to the problem of large overhead of image storage space.However,the general image compression methods don’t consider the characteristics of tunnel images,and its compression ratio is not enough to significantly reduce the storage cost.Therefore,a tunnel image compression method with high compression ratio is studied in this paper.The main work and innovation are summarized as follows:First of all,considering the characteristics of tunnel images,we propose a high compression ratio of tunnel images based on sparse coding method.Meanwhile,taking into account the differences of tunnel image content in different patches,we design a self-adaptive coding algorithm based on above mentioned differences,that is,more coding coefficients are assigned to texture regions with rich image content,while less coding coefficients are assigned to regions with background.In addition,for the obtained sparse coding coefficients,a simple and efficient quantization method is designed to further improve the coding efficiency.In the sparse reconstruction stage,the learned dictionary is used,and then the sparse coefficients and the dictionary are combined to obtain the reconstructed image.Finally,compared with the existing compression methods,the proposed method achieves higher compression ratio and better reconstructed image quality,especially when the bit rate is less than 0.05 bpp,the proposed method enjoys better visual results and outperforms JPEG and JPEG2000 in terms of Peak-Signal-to-Noise Ratio and structural similarity.Besides,we also apply the deep learning method to the tunnel image compression,and propose a tunnel image compression method based on the variational autoencoder.Firstly,the tunnel image is fed to the encoding network in the coding stage,and then a series of convolutional neural networks are used to extract features,reducing the spatial resolution of the input tunnel image,and obtaining the representation of the tunnel image in the hidden space.And finally use the quantization technology to get the compressed data.In the decoding stage,we apply inverse quantization operation to the compressed data,and then the data after inverse quantization is fed to the decoder network.In contrast to the encoder networks,some upsample operations are used to the feature map to obtain the reconstructed image.Experimental results show that our proposed method is efficient and effective compared with other image compression method in lower bitrates.