Research On Object Detection Based On Convolution Neural Network

Object detection based on convolutional neural network is one of the most important technologies in computer vision task,which has a very high practical value in face recognition,automatic driving,medical image diagnosis and other fields.But it has encountered many bottlenecks in the development process.For example,the current mainstream object detection models make label feature map based on anchor and all set feature pyramid module,but the setting of anchor seriously affects the universality of the detection model,and feature pyramid module will lead to a large number of redundant operations of the model;at the same time,these mainstream object detection models can not effectively extract and apply the associated features between the objects,and there is no method used to deal with the instability of model training caused by the large difference of object scale.In view of the above problems,the main research contents of this paper are as follows:Firstly,aiming at the negative influence of anchor on the universality of detection model and the problem that redundant operation of the model will affect the training efficiency of the model,an anchor free object detection model based on single route convolution neural network is proposed.First of all,a method of making label feature map at pixel level is proposed.The model can detect the object without anchor,which can avoid the setting of super parameters related to anchor and has higher universality.Then,the hourglass backbone network is used to replace the feature pyramid module to avoid redundant and complex calculation when making tag feature map,so that the model can pass through a single path Finally,we can use higher resolution feature map for training,so that the model can maintain high accuracy without anchor.Secondly,in view of the problem that the correlation features between the object to be detected in the image are difficult to be extracted by the object detection model and the scale resolution of the object to be detected is too large,a multi-scale selective object detection algorithm based on the convolution neural network of correlation feature extraction is proposed.First of all,the extraction module of object association feature is constructed,which is added to the model explicitly to participate in the pre-training,so that the model can obtain the extraction ability of object association feature;then,after the pre-training,the extraction module of model association feature is removed,and all objects participate in the model at a similar optimal resolution through the reduction of input image and the selection oftraining object,training can effectively avoid the problem of data distribution changes between datasets,as well as the problem of low efficiency of model weight updating and unstable representation ability caused by scale changes.Finally,multi-scale selective prediction is used in the prediction stage to ensure that the input objects in the whole training and prediction process are close to the optimal performance scale of the model.The experimental results on the COCO dataset verify the validity of the proposed model.There are 36 figures,9 tables and 72 references in this thesis.