Research On Edge Detection Of Multimode Fiber Imaging Based On Deep Learning

In the analysis of endoscopic images in the medical field,edge detection can be used to measure and diagnose the size,shape and position of lesions,as well as to automatically detect and identify lesions.At present,multimode optical fiber is regarded as the most ideal optical device in the field of endoscopic imaging.Multimode optical fiber imaging can directly transmit images through various spatial modes,which avoids the problems of diameter increase caused by the increase of the number of optical fibers in optical fiber bundle imaging,and compressive sensing and other restoration algorithms can effectively reconstruct multimode fiber images with high definition with less sampling rate.At present,the traditional edge detection of multimode fiber imaging is to first get a sufficiently clear multimode fiber imaging,and then use the edge detection operators to perform edge detection on it.The ensuing problems are that it is difficult to ensure a clear image in the imaging stage,and the errors in the imaging stage will be amplified in the edge detection process.In this paper,we try to use the method of deep learning to solve the problems of low efficiency poor edge detection in the field of multimode fiber imaging.In general,we design a neural network that outputs edge detection results end to end.The value sequence of bucket detector in the process of multimode fiber detection can be directly input into the neural network,and the edge structure of the detected object can be directly output end to end.We fully analyze and research the performance of neural network algorithm in multimode fiber imaging environment.Both simulation and experiments show that the neural network algorithm is more suitable for the edge extraction field of multimode fiber than the traditional compressed sensing algorithm combined with Canny operator.This algorithm can not only obviously improve the performance of edge detection of multimode optical fiber image and reduce the operation steps,but also show good performance in the case of optical fiber interference.In the case of extremely low sampling rate less than 1%,this paper also proposes an algorithm using deep learning combined with compressive sensing.Simulation and experimental results show that the algorithm proposed in this paper still has superior performance in the case of extremely low sampling rate.