| Low light level?3L?image is a kind of image,which is detected under the environment of low illumination(less than 10-3lux).At present,it has played an important role in many fields such as night vision of military,security and protection,remote sensing,etc.Owing to the low illumination of the detection environment and the dark light,the 3L image inevitably has the problems of low brightness and low contrast,the target details in image are blocked when the noise distribution is serious,neither subjective evaluation nor algorithmic processing can further identify the target.In this paper,3L image is taken as the research object.Firstly,the 3L imaging system was built to achieve image acquisition.Then,in order to improve the visual effect of 3L image,relevant denoising and enhancement algorithms were studied.Finally,the target recognition and detection of 3L image was realized based on the deep learning theory.The main research contents of this paper are listed as follows:1. In order to obtain 3L image,photon counting imaging theory was studied,the model was built to calculate the probability and the number of detected target photons after reflection.The 3L imaging system with MPPC?Multi-Pixel Photon Counter?as the core detection device was designed and built on hardware,mainly composed of MPPC,stepper motor controller,two-dimensional electronic control rail,computer,optical fiber,etc.A photon number acquisition programming and a visual interface were wrote on software,which could display the number of photon's acquisition in real time.By controlling the stepper motor,the two-dimensional electronic control rail was moved to scan the measured target row by row,the3L image was obtained by high-precision sampling and inversion calculation to target light field from time and space.2. The 3L image has many problems such as low brightness and low contrast,noise,etc.So a denoising and enhanced solution was proposed based on block matching filter and NSCT in this paper.The original 3L image and the histogram-equalized 3L image were transformed by NSCT at the same time,and their high and low-frequency coefficients were fused.The low-frequency sub-band was processed by the SSR algorithm.The high-frequency sub-band with the best preserving line features was denoised by the symbol function and Bayes-Shrink threshold to achieve primary denoisng.About the use of block matching filter,the local search was used to replace the global search,significantly improving the efficiency of the algorithm and it has a good denoising performance too.The simulation results show that the 3L image processed by the proposed solution has a better retention of details.Compared with classic algorithms like BM3D,it has 2%-9%performance improvement on the evaluation of some objective indices,and it lays a foundation for the follow-up recognition research.3. Based on deep learning theory,an improved detection method of improved Faster R-CNN for night vehicle images was proposed.The multi-features fusion was used on the basis of VGG-16 network,and the feature pyramid was added,which effectively improved the generalization ability of the model and increased the resolution of feature mapping.K-Means++cluster algorithm was added into RPN,which could obtain the optimal size and number of anchors.The loss function was modified by dynamically scaling the cross entropy,which strengthened the difficult samples mining of night vehicle images.Experiments show that the proposed method based on Faster R-CNN effectively improve the detection recall rate and accuracy of targets such as vehicles,traffic lights compared with original Faster R-CNN method.Among them,the detection accuracy of vehicles is improved by 3.12%,and the problem of missing detection of small target vehicles is effectively solved. |
PDF Full Text Request