Weak supervision target positioning method and device based on shallow feature background suppression

Abstract

The invention discloses a weak supervision target positioning method and device based on shallow feature background suppression. The method comprises the following steps: acquiring an image data set; constructing a target positioning classification network based on a neural network, wherein the target positioning classification network comprises a classifier, a generator and an evaluator; the classifier generates a shallow feature map and initializes a generator to obtain a foreground activation map; performing point multiplication on the foreground activation image and the shallow feature image, inputting into an evaluator to obtain a classification result, and calculating foreground cross entropy loss of the classification result and an original label; performing background principal component analysis on each image in the training data set by using a PCA method, obtaining a background activation image through clustering, and inputting the background activation image into an evaluator to calculate background classification loss; foreground-background confrontation loss is introduced, a Transform auxiliary generator is used for learning, and a total loss function is obtained; and inputting a test data set to obtain a positioning result. According to the method, shallow feature information is utilized to suppress the background to train the target positioning network, classification and positioning tasks are unbound, and the limitation of classification on the positioning effect is avoided.

Description

technical field [0001] The invention belongs to the technical field of weakly supervised target localization in computer vision, and in particular relates to a weakly supervised target localization method and device based on shallow feature background suppression. Background technique [0002] Weakly-supervised target localization is a sub-problem under target localization. Compared with fully-supervised (given object frame and category) target localization, weak supervision only needs to use the more easily obtained annotation information as a supervision signal, such as image-level labels, to To achieve a better localization effect, how to use only the category label information to maximize the detection of the location information of the target in the image has become a major problem in current research. [0003] In the prior art, a localization learning method based on class activation map is used for target localization; in the document "A-fast-rcnn: Hard positive gener...

AI Technical Summary

Problems solved by technology

Although these multi-stage erasure methods have achieved promising performance, progressive erasure will cause adversarial learning to introduce background noise and consume a lot of memory resources, and is limited by the random loss of information regions, leading to classification degradation; while ignoring different Adversarial erasure and different unerased / erased results between images, making it difficult to complement each other to mine the entire object

Since sample correlation is an effective method to mine reliable localization effects, the researchers proposed an image pair feature correlation target localization learning method; the document "Mining cross-image semantics for weakly supervised semanticsegmentation" proposed an image pair Random and global consistency localization methods for communication to keep features of the same category close in the high-level feature space; the document "Show, match and segment: Joint weakly supervised learning of semantic matching and object co-segmentation" has introduced a method aimed at understanding object co-segmentation. A co-attention mechanism for patterns for cross-image common semantic mining within paired images; co-segmentation proposed by Xie et al. in "Online refinement of low-level feature based activation map for weakly supervised object localization", using a geometric model to determine Dense correspondence between two images; however, the above methods only mine the correspondence in image pairs, ignoring the category-level feature correlation during the whole training process, resulting in low performance of object localization

There is also an object localization method learned by a semantic information activation map generator, by designing a two-stage weakly supervised object localization framework, a low-level feature-based activation map generator is proposed in the first stage to explore the underlying information in low-level features; It mainly consists of an image classifier and an activation map generator with a classification head (referred to as the generator); the generator is directly incorporated into the shallow layer of the image classifier, and the online activation map based on low-level features is jointly supervised by two classification losses Generated; but in most cases, the pixel values ​​in the activation map tend to hover near the boundary area, and the final target positioning effect may also spread to the background area, so the positioning effect is low

Images