Object Detection From Image Based On Convolutional Neural Networks

Image object detection is the basic problem of computer vision information processing.It is the basis of other advanced and complex visual problems such as behavioral understanding,scene classification and video content retrieval.It has the dual significance of academic research and practical application.In this paper,a candidate region proposal generation method based on multi-level feature fusion region proposal networks is proposed to solve the complex background interference problem.The multi-scale convolutional neural network is studied to extract feature adaptively.Multi-task classification and detection regressor is developed to classify the targets effectively.The main work of the paper is as follows:On the candidate region extraction level,there are complex background and irrelevant information interference problems in the target image.The RPN based on multi-level feature fusion is used to extract the candidate region.Do max pooling operation to low-level features and deconvolution operation the high-level features to keep the underlying features and high-level features at the same resolution.And then through the local response to normalize the underlying features and high-level features compressed to the same feature space.Finally,to extract more detailed regional features and obtain a small number of fine candidate region proposals,the ROI pooling layer and the full connection layer are added befor the classification layer and the detection layer.It can eliminate the interference of complex background and irrelevant information.On the feature extraction level,the traditional artificial features are difficult to design.The distance of the objects in the image to be detected and the illumination condition are different.The paper constructs a multi-scale convolution neural network model by constructing different scale convolutions in the network.In the bottom of the convolution layer to set a smaller convolution kernel to extract more details of the underlying location information.In the high-level convolution layer to set a larger convolution kernel to detect the global features of category.Finally,the ROI pooling layer is connected to map the whole feature of the object to the same feature map.Then the multi-scale feature is constructed.It can obtain the adapitive complex abstract feature of the imageand solve the problem of multi-scale / illumination change and the difficulty of extracting the image object detection feature.At the level of classification detection,this paper proposes a multi-task classification and detection regressor,which combines feature extraction,object classification and detection task in the deep learning framework.It establishes a complex mapping between image object feature and object classification and detection mode mechanism.Then the non-maximal suppression method is used to obtain the object detection area with higher final score.It improves the classification accuracy and corrects the ability of the convolution neural network to extract the effective features and solves the problem of difficult objects classification.Based on the experimental results of PASCAL VOC2007 and VOC2012,the paper proposes an image target detection method based on convolution neural network,which can improve the detection accuracy and achieve the image target detection.