Research On Target Detection Based On Mask-RCNN

Target detection technology is an important branch of computer vision research,which is used in many fields.If the images contain overlapping objects,it is easy to cause missed detection of the obscured objects during the detection process.How to effectively detect the situation where there are many objects in the same image and there is overlap has always been a difficulty.In the actual training process,there are various pictures used for detection.Due to the large gap between the width and height ratio of the target in the picture,the accuracy of the traditional target detection model is not high.How to detect this type of picture has important research value.At the same time,how to obtain high-quality pictures required for experiments from a large number of data sets has always been a research hotspot.It is time-consuming and labor-intensive to clean based on manual data.How to use a computer to automatically clean data is an urgent problem to be solved.In view of the above problems,this paper draws on relevant algorithms to change the Mask-RCNN network model,and cleans the data to improve the detection accuracy.The specific research is as follows:(1)Improve the size of the network-RPN selection of anchor points in the Mask-RCNN area.For special pictures with different aspect ratios,in order to improve the model generalization ability,select anchor points of different sizes and set the anchor point aspect ratio to[1:3,1:1,3:1],to improve the accuracy of detecting images with an aspect ratio in this interval.(2)The Mask-RCNN model is improved to detect overlapped objects.In view of the problems of the original Mask-RCNN model in detecting the objects with high occlusion and overlap in the pictures,different optimization algorithms are analyzed.In this paper,the NMS improved algorithm based on convnet,the soft NMS improved algorithm and the IOU guided NMS improved algorithm are introduced,and the soft NMS is introduced to improve the original Mask-RCNN.(3)The data cleaning network cleans the data.At present,data cleaning generally adopts manual screening,and the screening efficiency is low.In this paper,the pictures are cleaned by deep learning.The cleaning data is divided into three parts,the first part trains the model,the second part is the data that needs to be cleaned,and the third part is the comparison data.The filtering process is divided into two stages.The first stage filters out pictures containing text and numbers,and deletes these pictures.The second stage filters pictures with low recognition rates,and deletes the pictures in the dataset that are not prominent and have low recognition rates.Through two stages of screening,the purpose of cleaning data is achieved.