Research On Image Retrieval And Classification Algorithm Based On Adaptive Multi-scale

Great amounts of data have developed tremendously under the era of big data.Large-scale image data appears in our life.How to effectively retrieve and classify large amounts of image data has become a hot research topic.Deep learning algorithms are widely used in these two fields due to their high accuracy and easy model building.However,existing algorithms still have some problems,such as insufficient utilization of extracted features and insufficient diversity of features.Aiming at the above problems,this paper proposes a series of methods to improve the image retrieval and classification accuracy,starting from improving the feature utilization and diversity.In order to make full use of the intermediate features of the network structure,this paper presents image retrieval method based on deep learning to hash and multi-space,which improves the utilization of features and the accuracy of image retrieval.Traditional image retrieval based on deep learning to hash only constrain the generation of hash codes in hash space by cross-entropy loss function.In this method,the similarity of images is constrained in both Euclidean space and hash space to generate a better binary hash encoding.Before the hash coding,the feature extraction of the image is constrained by the contrast loss function in the Euclidean space.Together,they can minimize the distance between similar images coding,and maximize the distance between dissimilar images coding.The hash coding can better reflect the characteristics of the image.Due to the obvious effect of intermediate features in image retrieval,an image retrieval method based on deep learning to hash and multi-scale features is proposed in this paper,which improves the utilization and diversity of features and improves the accuracy of image retrieval.The method effectively combines high-level and low-level distinguishing features and intensifies the difference distinguishability between similar and dissimilar images.Specifically,it selects distinguishing features from high-level and low-level features by L1-norm importance standard.Fuse them at a certain proportion for image retrieval.In order to improve the resolution between the images,the method uses the mixed loss function,that is,the KL-divergence loss function preserves the dissimilarity between the dissimilar images and maximizes the distance between the dissimilar images.The cross-entropy loss function preserves the similarity between similar images and minimizes the distance between similar images.In order to solve the problem that is very time consuming for finding the feature proportion by a lot of tentative experiments in the image retrieval method based on deep learning to hash and multi-scale features,an adaptive multi-scale image classification method is proposed.This method allows the neural network to select the multi-scale feature fusion proportions by itself,which improves the feature diversity and image classification accuracy,and consumes less time.It uses a convolution kernel equivalence mechanism,and combines the mask and Sigmoid function to pick out the appropriate size of each convolution kernel through training.Specifically,each convolution kernels should have one or more parameters to control the selection of the size of the convolution kernel.Combined with modified Sigmoid function,these parameters are trained to zero or one in a differentiable training way.Zero means not using its corresponding size,and one means using its corresponding size.The neural network can determines the proportion of convolution kernel of different sizes required in each layer by itself,and can find the proportion that cannot be set by manual design.According to the structural characteristics of neural network in different size datasets,this paper also uses a regularization method,which can use more small-scale convolution kernel in small datasets and more large-scale convolution kernel in large datasets.It further improves the accuracy of image classification.It also uses weight sharing.Some weight parameters of convolution kernel of different sizes are shared,which can greatly reduce the consumption of resources during training.