Research On Methods Of Image Classification And Object Detection Based On Deep Learning

Image classification and object detection are the key technologies of many artificial intelligence applications in computer vision,and are also important research directions.Among them,image classification is the problem of classifying objects in an image.And object detection is to realize the recognition and localization of multiple objects in an image.In recent years,researchers have greatly improved the performance of image classification and object detection by using deep learning method.Feature learning using convolutional neural network has become the main method of image processing.However,image classification and object detection are still very challenging due to the change of the scene,object occlusion,ambiguity objects and low-resolution images,etc.In view of the above problems,this paper is based on deep learning technology,to take the acquisition of effective features as the breakthrough point,and to analysis and discuss image classification and object detection methods from different aspects.The main contents and innovations of this paper are as follows:(1)A novel bi-branch deconvolution-based convolutional neural network is proposed and achieved good image classification performance.Convolutional neural networks usually use convolution for feature learning,but this way is relatively simple,and the learned features are not rich enough to affect the final recognition performance.To this end,this paper proposes a bi-branch deconvolution-based convolutional neural network.Using deconvolution to learn features with high semantic information and edge information,and combining with the features of convolution learning,a feature extraction module with two branches is constructed.The two branches use different scale convolution filters to capture different features of the object.Then the image classification model is constructed by simply stacking the module.In this paper,the method is verified to improve the identifiability of features and obtain better target recognition performance on three image classification data sets.(2)A novel one-stage object detection model based on densely convolution and feature fusion is proposed.The model achieves better target detection performance without the pre-training model.At present,most of the object detection models are based on the pre-training model on Image Net,which limits the extensibility and generalization of themodels.In this paper,a one-stage object detection model without pre-training is proposed.The main frame of the model is composed of several densely convolutional blocks to extract the powerful multi-scale features.Stacking multiple densely convolutional blocks can overcome the gradient dissipation problem that may occur when the network reaches a certain depth.Aiming at the problem of feature loss caused by one-way transmission of information flow in the model,especially the problem that the detail features and the contour features are lost with the deepening of the network,a feature fusion module is introduced into the model.It enriches the object features by separately processing three adjacent scale features into the same scale and then effectively fusing them to achieve better object detection performance.Experiments on PASCAL VOC and MSCOCO object detection data sets show that the model without pre-training can achieve similar object detection performance as other state-of-the-art object detectors.(3)A novel one-stage object detection model based on rich global context is proposed.During the object detection process,all objects are detected based on the local areas.Adding global context information to the features can expand the receptive field to reduce uncertainty in local areas and increase detection accuracy.In this paper,a learning model of object detection based on rich global context information is proposed.The global activation module is added to the main frame to emphasize the integrity of the features and weaken the features with less effect in the local areas.At the same time,a pyramid feature pool module is constructed to generate multi-scale global context information,which is used to supervise the multi-scale object features extracted from the main framework,so as to enrich and discriminate the final target features,thereby improving the accuracy of object detection.Experiments on PASCAL VOC and MSCOCO object detection data sets show that the model performs well without relying on the pre-training model,especially in the detection of occluded objects and small objects.