Pooling method based on bilinearity and spatial pyramid

Abstract

The invention discloses a pooling method based on bilinearity and a spatial pyramid, and belongs to the field of image processing and computer vision. The method comprises the following steps: acquiring a video stream, and intercepting a to-be-processed target image; extracting features of different levels or different categories in the target image; fusing the feature groups through a bilinear method to obtain a global feature map; carrying out pyramid pooling on the fused global feature map to reduce the dimension of the feature map; and normalizing the feature map after dimension reduction, taking the feature map as a final feature of the target image, completing pooling operation, and applying the obtained final feature to subsequent classification to realize identification of the to-be-detected object. The invention is suitable for behavior recognition in images and pooling operation in target detection, the dimension of multi-feature fusion is reduced, the recognition efficiency is improved, and different recognition requirements for multiple features in recognition are met.

Description

technical field [0001] The invention relates to the fields of image processing and computer vision, in particular to a pooling method based on bilinear and space pyramid. Background technique [0002] In the era of rapid development of intelligent science and technology, the functions of behavior recognition and target detection of intelligent monitoring are gradually improved and popularized, and the convolutional neural network often uses pooling operations to reduce the dimension of the feature vector output by the convolutional layer. Improve results with minimal impact on expressive semantics of the original image. Due to the "static" characteristics of images, in different image regions, useful features can often be shared and applied. It aims to imitate the human visual system, and the pooling operation can aggregate and count features at different locations. [0003] Traditional pooling methods generally include average pooling, maximum pooling, and random pooling, ...

AI Technical Summary

Problems solved by technology

[0005] In order to solve the defects of the existing image pooling methods for behavior recognition, object detection, etc., the present invention combines bilinear pooling and pyramid pooling. First, multi-feature extraction is performed on objects in the target image, and the feature group Perform bilinear fusion to obtain the fused global feature map, and then perform pyramid pooling on its corresponding position

Images