Research And Application Of Deep Neural Network In Image Recognition

Computer vision is currently the most significant direction in the field of artificial intelligence,which aims to teach the computer visualizing the world via images.As a main focus of computer vision,image recognition shows its own distinctive superiority in surveillance,biometrics,unmanned driving,medical diagnostics and etc.More recently,the application of deep learning algorithms to image recognition is a grand achievement which results in the continuously improving of recognition accuracy.Whereas,the deep learning based recognition models still have to tackling the problems of sample collecting conditions,semantic analysis,computing complexity,model robustness and insufficient training samples.For this reason,this paper primarily concentrates on improving the involving techniques of deep neural networks for image recognition.The innovative contributions are summarized as follows:1.The widely-used feature pooling(aggregation)approach can cause the issue of over-fitting,and further lead to the poor robustness of the model.This paper propose the a method of weighted pooling,which is based on establishing the feature map of each local area using the information entropy.Specifically,the mutual information is taken to measure the significance of each feature within each local area.Experimental results indicate that the weighted pooling can not only precisely compute the activation representative of the local area,but also effectively improve the recognition accuracy and generalization ability of the model.Notably,compared to the classical feature pooling methods,the experiment on CIFAR-10 test set verifies its capability as the maximum rise of the recognition rate reaches 2.7%.2.Generally,the deep learning models have the shortages of slow convergence speed,gradient easily disappearing and falling into local optimum.As such,these difficulties are highlighted during the training process.In this paper,the algorithm of Adaptive multipoint moment estimation(AMME)is proposed by using the multi-point moment estimation and weight attenuation.On the one hand,each moment estimation point is taken to inspect the skewness and kurtosis of the error gradient,which further speeds up the parameter updating during model training.On the other hand,the employing of weight attenuation method can thus enhances the model robustness.Testing on datasets of MNIST,CIFAR-10 and etc.shows that the AMME outperforms other model training methods in both convergence speed and recognition accuracy.Distinctively,the improvement in convergence speed is 3%and that in accuracy is 1.1%in comparison to the baseline algorithms.3.Image datasets are crucial to the deep learning models.The training of models requires large amount of samples while the costs of detection samples are quite expensive.Thereby,by revising the pixel weight calculating approach,an unsupervised learning method is presented to transfer the classification samples to detection samples.In this way,the issues of high cost,insufficiency of sample size and variety of the detection dataset is addressed.Besides,current detection models mainly relies on the experience of researchers since the shape and number of the bounding boxes are set manually.However,this strategy is absent of objective basis and adaptability.On this occasion,by combining the properties of a given dataset,the distance calculation is revised within the k-means++clustering.Furthermore,the shapes of the objects in training samples are clustered to obtain the prior knowledge of the bounding boxes.Accordingly,the detection rate and convergence speed are rised.Lastly,based on the prior distribution of the object together with the aforementioned improvement,the Statistic Experience-based Adaptive One-shot Detector(EAO)is developed.Experiments are conducted on the dataset of PASCLE VOC and MS COCO.Clearly,comparing to the state-of-the-arts,our model is a better alternative by increasing the accuracy by 0.6%and 1.5%on the two datasets,respectively.In summary,this paper focuses on three key issues in image recognition technology.In terms of model components,a weighted feature pooling method and an adaptive multi-moment moment estimation stochastic optimization algorithm are proposed.In terms of model,an adaptive single-network target detection model is proposed.The experimental results based on the public dataset show that the proposed method achieves good results in image recognition tasks,and relevant theoretical results have been published in the journals of artificial intelligence.