Study On Object Detection And Recognition Methods For Complex Scene Images

As an important branch in computer vision,the main task of object detection and recognition is using image perception models and algorithms to enable computers to find objects of interest from images,while automatically classifying and locating these objects of interest.With the development of deep learning,object detection and recognition models based on deep learning relying on a large number of labeled samples and sufficient computing resources show an excellent performance in accuracy and speed,and widely used in automatic driving,urban transportation,smart medical treatment,industrial detection,environmental monitoring and other fields.However,obtaining and labeling samples of object detection and recognition usually takes a lot of manpower and material resources.In addition,due to the influence of different lighting,weather,shooting background,shooting angles and shooting distance,objects in the collected images are usually multi-scale,deformed,easily blurred,etc,which greatly increases the difficulty of object detection and recognition.Therefore,from the perspectives of sample expansion and model construction,this dissertation studies the sample generation,sample selection,feature enhancement,and difficult sample learning in object detection and recognition to improve the accuracy and generalization of object detection and recognition in complex scenes.The main innovations are as follows:(1)Aiming at the problem of small number of samples and poor diversity,diversity sample augmentation methods based on generative adversarial network is studied,and we propose a sample augmentation method based on multi-branch conditional generative adversarial network.This proposed method designs and trains a multi-branch conditional generative adversarial network for generating multifarious object images.To ensure the quality of sample augmentation,a selection method for generative samples is designed to select objects images that are conducive to model training.In addition,in the process of sample augmentation,according to the proportion of the object number of different categories in the original training sample set,the generated objects are inserted into the training sample set evenly to maintain the balance between the number of different categories of objects in the training sample set,which effectively improves the training effect of the object detection and recognition model.Finally,the object generation experiment and sample augmentation experiment show that sample augmentation method based on multi-branch conditional generative adversarial network can effectively improve the accuracy of object detection and recognition models.(2)Aiming at the high cost of manual sample labeling and the lack of high-value samples,sample augmentation methods based on active learning is studied,and we propose a novel active learning method based on global-margin uncertainty and collaborative sampling.This proposed method firstly calculates the global uncertainty and marginal uncertainty of objects according to the scores of the predicted objects on different categories,then combines the two uncertainty as global-margin uncertainty to estimate classificatory uncertainty of objects.To effectively utilize the prediction information of object location,a collaborative sampling query strategy is designed.On the basis of global-margin uncertainty,the difference between the prediction results of two detectors is used to further estimate the uncertainty of unlabeled images,then the images with high uncertainty are selected as high-value samples for labeling,which effectively reduces the cost of manual labeling and improves the quality of sample expansion.Experiments on two datasets show that the proposed active learning method based on global-margin uncertainty and collaborative sampling can effectively reduce the cost of manual annotation,enhance the quality of training samples,and improve the training effect of object detection and recognition models.(3)Aiming at the problem of poor discrimination of object feature,object detection and recognition methods based on feature fusion is studied,and we proposed a novel object detection and recognition method based on multi-scale feature fusion.Based on the feature pyramid network,this proposed method designs a feature reinforcement component,which uses residual block and transpose convolution to enhance the semantics of underlying feature maps,and generate the enhanced feature maps for feature fusion.In addition,a hard fusion strategy and a soft fusion strategy are also designed to fuse the underlying feature maps,high-level feature maps,and enhanced feature maps to generate feature pyramids,making the features of different scales can be better integrated and matched,which effectively enhances the discrimination of object features,and improves the accuracy of object detection and recognition in different scenes.Finally,the experimental results on four datasets show that the object detection and recognition method based on multi-scale feature fusion is superior to several advanced object detection and recognition methods in accuracy.(4)Aiming at the problems of large intra-class difference and low identification of objects in complex scenes.Object detection and recognition methods based on multi-level feature and metric learning is studied,and we propose spatial hierarchy feature perception and hard samples metric learning for object detection and recognition method.In this proposed method,the convolutional layers of different receptive fields are utilized to extract different spatial hierarchical features,and a weight vector is used to learn the weight between feature channels,which effectively enhances the feature representation.We also designs a hard sample metric learning method to select the samples with large loss value as the hard samples,and close the distance between the features of the hard samples in the same class through metric learning,which reducing the false and missed detection caused by the intra-class difference of the objects.In addition,a pre-training set is established to pre-train the object detection and recognition models,making the model learn the object features fully,which further improving the accuracy of object detection and recognition.Experiments on two datasets show that,the proposed object detection and recognition method based on method spatial hierarchy feature perception and hard samples metric learning has excellent performance in accuracy,compared with more than ten advanced target detection and recognition methods.This dissertation has 65 figures,28 tables and 179 references.