Object detection method using feature image fusion

Abstract

The invention discloses a target detection method using feature image fusion, which comprises the following steps: firstly, extracting image features by using a ZF network and obtaining a series of feature images at different levels; secondly, obtaining the new feature map by fusing the deepest feature map and the shallow feature map extracted by ZF network; inputting the new feature map into RPNnetwork again to get the area suggestion; finally, the new feature map and the region suggestion are input into the ROIPooling layer to get the region suggestion features, and the classification of the features and the border regression of the region suggestion are used to get the target detection results. The invention can detect a plurality of kinds of targets in an image, and the image used does not need a specific image acquisition device to acquire.

Description

technical field [0001] The invention relates to the field of computer vision, in particular to an object detection method utilizing feature map fusion. Background technique [0002] Target detection has always been one of the important and difficult topics in the fields of event recognition, intelligent transportation, etc. Its task is to locate and classify a variable number of targets in the image to be detected. The target position is marked in the form of a box, and the classification is to determine the category of the target in the image. [0003] The existing target detection methods can be divided into two categories: traditional target detection methods and deep learning-based target detection methods. Due to the gap between the detection effect of traditional methods and the actual needs in some cases, the target detection method based on deep learning uses convolutional neural network (CNN) to automatically extract image features, and the extracted features are m...

AI Technical Summary

Problems solved by technology

However, most of the existing object detection methods based on deep learning only use the deepest feature map in the feature map extracted by CNN when using CNN to extract image features, and do not use the shallow feature map, and the target detection accuracy Cannot meet actual needs in some respects

The method proposed by Girshick R and Donahue J in the article "Rich feature hierarchies for accurate object detection and semantic segmentation. Computer Vision and Pattern Recognition. IEEE, 2014:580-587" has the problem of computational redundancy in the process of using CNN to extract region proposal features , and the target detection accuracy of the proposed method is relatively low

The method proposed by Girshick R in the article "Fast R-CNN.Computer Vision.IEEE,2015:1440-1448" is non-end-to-end, and the target detection accuracy is relatively low

The method proposed by Redmon J and Divvala S in the article "You only look once: Unified, real-time object detection. Computer Vision and Pattern Recognition. IEEE, 2016: 779-788" has relatively low detection accuracy for small targets

Images