Fine-grained recognition method of image

Abstract

The invention discloses an image fine-grained recognition method, which specifically comprises the following steps that: according to image data, obtaining an original picture set and a disturbing picture set, the disturbing picture set is formed by randomly disturbing processed pictures in the image data, and recording a random disturbing sequence corresponding to the pictures; respectively inputting the original picture set and the disturbing picture set into a deep neural network to obtain an original feature map and a disturbing feature map; restoring the disturbing feature map into a newdisturbing feature map according to the random disturbing sequence; performing pooling operation on the original feature map and the new disturbance feature map, inputting the original feature map andthe new disturbance feature map into a softmax classifier for classification to obtain an original loss function and a disturbance loss function, and calculating an Euclidean distance between the original loss function and the disturbance loss function; and calculating a total loss function according to the original loss function, the disturbance loss function and the Euclidean distance between the original loss function and the disturbance loss function. An input image is divided into a plurality of local areas, and a confusion and restoration method is adopted, so that a network learns local area features and semantic information, and a fine-grained recognition task can be completed.

Description

technical field [0001] The present invention relates to the technical field of deep learning, in particular to an image fine-grained recognition method. Background technique [0002] The process of deep learning for classification and recognition tasks is usually: target detection, position correction of the detected target, and then sent to the neural network to extract features, and finally classified. In existing technologies, classification is usually based on global features. The disadvantage of classifying global features is that some subtle features may be ignored, some features unique to a certain category. For example, it is necessary to identify the uniqueness of the car. The only difference between two identical Audi cars is the position where the annual inspection label is pasted. If global features are used for representation, the subtle feature of the annual inspection label may be ignored. Therefore, With fine-grained image recognition tasks, it is mainly us...

AI Technical Summary

Problems solved by technology

[0007] The present invention proposes a fine-grained image recognition method, which solves the need for labeling information on objects or local areas in the prior art, which requires a lot of labor c

Images