Image deep characteristic determination method and system

Abstract

The invention discloses an image deep characteristic determination method and system. The method comprises the steps of extracting a characteristic image output by each convolution layer in the deep convolution nerve network; calculating the layer entropy and the relative layer entropy of the convolution layer; selecting a characteristic expression layer from the convolution layer based on the layer entropy and the relative layer entropy of the convolution layer; establishing a deep characteristic descriptor based on the characteristic image output by the characteristic expression layer; conducting coding on the deep characteristic descriptor, and obtaining the deep characteristic expression of the input image. The invention is advantageous in that generalization is strong, processing is simple, and high-rise characteristic expression having robustness and discrimination can be obtained.

Description

technical field [0001] The present invention relates to the technical field of image analysis, in particular to a method and system for determining image depth features. Background technique [0002] How to obtain robust feature representation has always been the focus of computer vision tasks. Traditional manual features are designed according to specific tasks and are suitable for fixed scenarios. When migrating, manual features must be designed according to the characteristics of the image. With different application scenarios, corresponding changes are made to the model design, and the generalization ability is poor. And the amount of calculation is large. Therefore, it is difficult to guarantee the accuracy of directly using manual features to express the target. [0003] Deep learning, by simulating the hierarchical structure of the human brain's visual perception system, establishes a machine learning model with a rich hidden layer structure, and can learn the essent...

AI Technical Summary

Problems solved by technology

However, DCNNs are computationally intensive and time-consuming during training and testing, and need to collect and label a large amount of data during training. Therefore, scholars have begun to pay attention to directly extracting features from pre-trained models.

The study found that the convolutional layer can extract local features from the target, but due to multi-level abstraction and compression, these features are difficult to preserve in the fully-connected layer (fully-connected layer, fc layer), while the fc layer can retain the global features and semantics of the image information, but poor geometric invariance

The existing feature layer aggregation method only selects the most suitable feature layer through the experimental results, and aggregates the feature map of this layer to form a feature descriptor, which is cumbersome and computationally intensive.

Images