Picture property detection method based on regional sensitivity score atlas and multi-instance learning

Abstract

The invention discloses a picture property detection method based on a regional sensitivity score atlas and multi-instance learning. The method comprises the following steps: converting an inputted image into an RSSM feature atlas by using a convolutional neural network; transforming the RSSM feature atlas into a MIL feature atlas with the dimension of 1000*10*10 by using an RSSM composition layer; and then inputting the MIL feature atlas into a multi-instance learning MIL network layer to obtain an attribute probability vector with the dimension of 1000*1. Compared with the previous methods,the picture property detection method disclosed by the invention enables the detection accuracy to be increased obviously. Under the same circumstance, the methods based on a CNN model and an FCN-MILmodel only have the accuracy reaching 30.8% and 34.0% respectively; however, the picture property detection method disclosed by the invention has the accuracy reaching 42.1%. Besides, the picture property detection method disclosed by the invention is capable of detecting 1000 kinds of attributes of the picture and thus has a comprehensive detection range larger than that of the common attribute detection method; and needs of common picture video description and scene understanding can be met basically.

Description

technical field [0001] The invention designs a picture attribute detection method based on region-sensitive score map and multi-instance learning, and relates to the technical fields of deep learning and computer vision. Background technique [0002] At present, object recognition, image classification and other problems have achieved good results with the rapid development of deep learning. These two types of problems can be classified as single / multi-label classification problems, that is, each target picture has a certain number of labels. The ultimate goal of class problems is to get these labels correctly. These labels all have one thing in common, that is, they belong to nouns and other words that can be substantiated. In the problem of image semantic information acquisition, there is actually a need to obtain the objects contained in the image and the vocabulary related to the object, such as verbs, adjectives, and nouns. Strictly speaking, it is also a multi-label cl...

AI Technical Summary

Problems solved by technology

[0005] Among the above methods, the methods based on KNN and Ranking belong to the traditional method, and the accuracy rate is not high; while the image attribute detection method based on CNN adopts the method of deep learning, but only considers the global characteristics of the image. In many cases, it only occupies a small part of the picture, and the same part of the picture may contain multiple attributes, which reduces the effectiveness of the CNN model for image attribute detection to a certain extent; FCN-MIL-based image attribute detection method It has the feature of identifying blocks of pictures, so that the attribute detection method can pay attention to the local area features of the picture while paying attention to the overall situation of the picture, and make full use of the information contained in the picture, so the accuracy rate is much improved compared with the CNN-based method. However, because the algorithm itself has the characteristics of block recognition, the size of the original picture associated with each value on the sub-map is directly related to the size of the MIL feature map. Therefore, if the size of the feature map is too small, the original image size associated with each value on it The image size will be too large, which will affect the feature extraction effect of the algorithm on the local area of ​​the image.

Therefore, the attribute detection effects of these methods are not satisfactory.

Images