The Research Of Object Detection Model Based On Convolution Neural Network

Object detection technology is the basic research direction in the field of computer vision,providing basic image information data for advanced computer vision processing and analysis tasks.With the continuous breakthrough of deep machine learning,especially the deep convolutional neural network model shows powerful image feature extraction capabilities in the field of digital image processing.Although the research on the object detection model based on CNN has developed rapidly,there are still some problems in the practical application.Such as the number of parameters of the detection model is large,high storage costs and computational costs are required.Region proposal-based object detection model with high detection precision have slow detection speed and are difficult to meet real-time requirements.Although the bounding-box regression object detection model can provide real-time detection speed,but its detection precision does not achieve the desired results,especially in the performance of small-size object detection.In this thesis,we make the following researches on the problems existing in the current CNN based object detection model:(1)There is a lot of redundancy in the large number of parameters of the CNN network model.In this thesis,the number of parameters of the CNN network model is compressed by using a depth-wise separable convolutional layer to replace the standard convolutional layer used in the traditional CNN model.Depth-wise separable convolutional layer split the extraction and the compression of image feature in two steps.Decompose a standard convolutional layer into deep-wise convolution layer and point-wise convolution layer to reduce the number of parameters.The results of comparative experiments show that using the depth-wise separable convolution layer instead of the standard convolution layer can greatly compressed the number of parameters of the CNN model,and the feature representation ability of the CNN model is reduced only within an acceptable range.In this thesis,the number of parameters of the Darkent-19 model convolutional layer is reduced by 78.83%,by introducing a deep-wise separable convolution layer instead of a standard convolution layer.(2)The object detection model based on the convolutional neural network does not make full use of the image features extracted from the based CNN network,which results in poor detection performance of small-size objects.The Feature Pyramid Network was chosen to fuse the image features extracted from each layer of the CNN network.The image features containing different information are fused,and the pyramid structure of the multi-scale image features is provided for the classification and localization of the object detection model.The object detection model can be used to detect the object of different size using the proper fusion image feature.The comparison experiments show that the introduction of the Feature Pyramid Network provides high model detection precision.The mean average precision on the PASCAL VOC dataset increased to 77.31%.