Image fine-grained recognition method based on multi-scale feature fusion

Abstract

The invention discloses a retinal vessel segmentation method based on dense convolution and depth separable convolution. The method comprises the following steps: preprocessing an original image of aretinal vessel map; carrying out data enhancement on the data set; constructing a full convolutional neural network based on combination of dense convolution and depth separable convolution, and training the training set by using a loss function with weight; and testing and obtaining a final segmentation result graph. Encoding-decoding symmetric network is used as a backbone network. The additiondepth can separate convolution, so that the model parameter quantity is greatly reduced; dense convolution blocks are used, all layers are connected on a channel, information transmission is enhanced,and characteristic values of all scales are effectively utilized; in the image preprocessing process, self-adaptive gamma correction is carried out on the image, different gamma values are used for correcting different feature areas, and background noise is weakened while the feature contrast ratio is improved; during training, a loss function with weight is used to enhance the proportion of theblood vessel to be segmented during training.

Description

technical field [0001] The invention belongs to the fields of computer vision, artificial intelligence, and multimedia signal processing, and in particular relates to an image fine-grained recognition method based on multi-scale feature fusion. Background technique [0002] Image fine-grained recognition is a subject of research on image subclass recognition and classification. With the rapid development of deep learning and artificial intelligence technology, the problem of image fine-grained recognition as a basic computer vision subject has also made great progress. The problem of fine-grained recognition is relative to coarse-grained recognition. Coarse-grained recognition refers to the recognition and classification of large categories of objects in the traditional sense. Fine-grained recognition refers to the identification of sub-categories, such as the CUB200-2011 data proposed by the California Institute of Technology. Concentrated identification of 200 species of ...

AI Technical Summary

Problems solved by technology

[0004] For the fine-grained recognition tasks of image target subclasses, some existing methods fail to combine the local features of the bottom layer well, and there are situations where the real-time requirements are poorly met. The present invention is based on the bilinear convolutional neural network. Combined with the idea of ​​multi-scale feature pyramid, the present invention provides a fine-grained image recognition method based on multi-scale feature fusion, which is characterized in that: the batch images after data augmentation are sent to the basic network to extract features, and the feature pyramid method is used to fuse multiple Scale features, bilinear features are extracted at each layer for independent prediction, and the prediction results of each layer are voted by the minority to obey the majority to obtain the final prediction result. The method includes the following steps:

Images