Research And Realization Of Pivotal Techniques Of Small Target Detection In Medical Image

Object detection has always been the basic task in the field of computer vision.The detection method for general target can not effectively solve the problem of small target detection.For the detection of small targets,the detection effect of small targets is not good due to the features of low resolution,fuzzy image and little information.At the same time,due to the scarce available data and complex detection background in the field of medical image,it is more difficult to detect the target than in other fields.In this paper,the target detection method based on convolutional neural network is studied and analyzed.On the basis of Faster R-CNN(Faster Region-Convolutional Neural Networks),an improved small target detection method for medical image is proposed.This paper focuses on the detection of small targets in medical images and conducts the following research:1.The medical image was preprocessed.Data preprocessing operations including image format conversion,image segmentation,image data enhancement and image screening were used to preprocess the collected medical images.Medical image preprocessing is an important part of medical image detection,and deep learning algorithm is an algorithm that relies on and requires a lot of data.Although there are some medical images on the public data set,this part of data cannot be directly used for the training of deep learning model due to the format and formation mode.2.A small target location method based on RPN(Region Proposal Network)improvement is proposed.Optimized GIoU(Generalized Intersection over Union)is used to replace the original smoothL1 as the loss function of RPN regression branch.Faster R-CNN in extracting the candidate region is RPN,used in analysis of regression process PRN found using IoU(Intersection over Union)as loss function is better than the original loss function smoothL1,but due to loss of ious directly as a function will have serious problems,and the emergence of GIoU.just to compensate for possible problems,after verification,Optimized GIoU as a branch of RPN return loss function better than directly using the IoU is better than the original loss function smoothL1,So we replace the loss function of the RPN regression branch with Optimized GIoU.3.A small target detection method based on improved Faster R-CNN was proposed.The method of multi-scale feature map is used to improve the recognition network,andthe feature map obtained by convolution several times before is spliced together after ROI pooling,so as to improve the resolution of feature map of small target area.Then input to the recognition network,object recognition.The problem that the recognition rate of Faster R-CNN on small targets is not high or even impossible to be recognized is effectively solved.The final comparative test shows that the improved small-target detection method proposed in this paper has achieved good detection results in precision and recall.At the same time,the design and implementation of medical image detection management system is carried out with the small target detection algorithm as the core.