Detection And Classification Of HNS On Water Surface Based On Faster R-CNN And Multispectral Imaging

HNS spill accident poses a huge threat to the ecological environment and public safety.Because most of HNS are colorless,the difference between HNS and the water background is small and the difference between different HNS is smaller.It is difficult to detect HNS by RGB images.In this paper,Faster R-CNN combined with the multi-spectral image classification technology was applied to detection and classification of floating HNS,which provide rapid technique for the emergency of HNS spill accident.In this paper,the reflection spectra and image characteristics of three kinds of transparent HNS floating on the water surface,such as benzene,xylene and vegetable oil,are studied.Spectral analysis,spectral image data extraction,and image band comparison of acquired reflection curves(325-900 nm)and multispectral images(365,410,450,500,550,600,650,700,and 850 nm)were applied After analysis,the results show that the spectral reflectance difference between the HNS sample and the water background was the largest in the ultraviolet wavelength.In the spectral image,the contrast between the HNS region and the water surface was larger in the 365 nm ultraviolet wavelength than in the visible light wavelength,which is more conducive to target recognition and Extraction;the spectral and image characteristics of different samples in each wavelengths were different,indicating that the multi-spectral band data has the potential to distinguish floating HNS samples.Further,for fast and automatic detection of the target area,the Faster R-CNN algorithm was applied to the RGB and UV(365 nm)images of the floating HNS.The results show that the average target frame overlap(IOU)between the predicted area and the real area of the UV image detection model was 0.869,slightly higher than RGB(0.855).The average classification accuracy of the UV model was 86.66%,which was significantly better than the 68.33%of the RGB model,shown that UV images were more suitable for target detection of floating HNS than traditional RGB images.In order to improve the classification result of the Faster R-CNN target detection algorithm,the multispectral images of the floating HNS region were segmented to extract data,and classification models based on multispectral image data were established.The results show that the LS-SVM model using only the four wavelengths(365,410,450 and 850 nm)can obtained the classification accuracy of 100%,which can effectively improve the accuracy of Faster R-CNN classification model.Therefore,Faster R-CNN combined with multi-spectral image detection technology can effectively detect transparent HNS.