High Quality Object Detection Algorithm Based On Convolutional Neural Network

In the context of the increasing influence of computer vision,target detection,as the basic research in computer vision,plays a very important role in the fields of face recognition,driver lessness and security.Among the commonly utilized target detection algorithms,the architecture based on region recommendation performs best in detection accuracy.It divides the network into three parts,which are the basic network,the recommended area generation network and the detection network.And the Convolutional Neural Network(CNN)is the infrastructure of these three parts.Although CNN is playing an important role,there are still key issues needed to be solved urgently.The CNN introduces some strong priors to balance large parameters and actual costs,which results in the decrease in positional accuracy as the network becomes deeper.Therefore,the CNN-based target detection network needs to be improved.This paper refers to the architecture based on regional recommendations to improve the basic network and detection network.The specific research content of this paper is as follows:(1)This paper proposes a high-precision target detection network IEPC-Net,which consists of the location information enhancement network and the cascading hybrid detection network.(2)Aim at the problem of location information loss of CNN,this paper proposes the location information enhancement network.The network adds a location-sensitive ROI pooling layer to the feature pyramid network FPN,Therefore,it can not only effectively utilize the precise position information of shallow features and the highly abstract information of deep features,but also avoid the loss of position information caused by the traditional ROI pooling layer.(3)Aiming at the problem of traditional detector and sample mismatch,this paper proposes a parallel cascade detection network.Although the existing cascade method can avoid detecting the network optimal adaptation interval and the sample quality distribution mismatch,the network is over-fitting due to multiple single detector cascades.The proposed method adds parallel detectors to the cascaded network,which improves sample quality and maintain sample diversity.Therefore,the method can avoid the mismatch between the detector and the sample.What's more,it can avoid the over-fitting problem of the network,which will effectively improve the performance of the detection network.