Image Classification And Object Detection Based On Convolutional Neural Network

With the rapid development of network and information technology,various media datas have grown rapidly.A large amount of visual data has emerged.Faced with various images and videos,how to identify users' interest data effectively is a hot problem,because various objects have different sizes,shapes,and always with occlusions,image classification and object detection have always been one of the most challenging problems in the field of computer vision.Image classification is to identify the image according to the contents of the image,determine which objects exist in the image,and assign a set of labels to the image.Object detection is to locate the interested object rom the image,and to locate each object.Classify and give a bounding box for each object.So far,many novel image classification and object detection methods have been born.Based on the predecessors,this paper further explores image classification and target detection techniques.Firstly,the paper analyzes the research background of image recognition and the research status at home and abroad.Then it discusses the key technologies of image recognition and focuses on several main work of the thesis:Chapter 3 proposes a multi-label image classification method based on adversarial network,which adopts adversarial network to improve the recognition accuracy of objects with occlusions.The fourth chapter proposes the object detection method based on multi-feature fusion.Finally,the main research work of the thesis is summarized in the fifth chapter and the future work prospects are expounded.The main research contents of the thesis are:(1)An image classification method based on adversarial network is proposed.Firstly,the VGG16 convolutional neural network model is combined with the spatial pyramid pooling to obtain more spatial feature representation of the image.Then the sigmoid cross entroy loss multi-label classification function is used instead of the single-label function to train and optimize the multi-label image.Finally,an adversarial network is trained for the joint optimization with the original network,the adversarial Spatial Droput Network(ASDN)is usedto deal with the features extracted by the original model,the window map of the corresponding slider sliding is deleted to generate the training samples with occlusion features,and the network is continuously trained with the difficult samples with occlusion,so that the network also has better recognition effect to occlusion objects.(2)A multi-feature fusion object detection method is proposed.Firstly,a cascade network model is established.The feature map structure with different convolution kernel sizes is used to extract multiple features of the image,so that the network can extract spatial information of different objects in the image.Secondly,this paper cites and improves the multi-scale pooling to perform multi-scale feature extraction on the feature maps,and introduces ROIAlign to improve the pooling,and solves the problem of regional mismatch caused by two quantization in ROI Pooling operation.This allows the model to extract small objects and objects of different shapes efficiently.Finally,in order to ensure the speed of object detection,this paper uses 1 × 1convolution to reduce the dimension of some features,avoiding the parameter redundancy and speeding up the object detection speed and model iteration speed.