Target detection and recognition method for image sequence

Abstract

The invention belongs to the technical field of image detection and recognition, and can be used in the fields related to image sequence processing, such as automatic piloting, detection and recognition of visual objects of unmanned aerial vehicles, and detection of animation video. Firstly, the input image is transformed by Image Warping (Image Warping) based on motion compensation, then the depth convolution feature maps are extracted from the winding images by depth convolution neural network, and then the depth convolution feature maps of the image sequences are fused. Finally, the fused depth convolution feature maps are used in target detection modules such as target location, target recognition or region recommendation. A main technical feature of that invention comprises two partsof a computing module of a winding image based on motion compensation in the early stage and a fusion module of a depth convolution feature map of a front image sequence and a rear image sequence in the late stage. The invention can overcome the problem that the detection accuracy of a single image is low or even cannot be detected due to partial occlusion, motion blur and large deformation of thedetection target.

Description

technical field [0001] The invention belongs to the technical field of image detection and recognition, and in particular relates to a detection and recognition method for an image sequence target. Background technique [0002] In the field of image understanding, deep learning technology is widely used in image classification, image object detection, face recognition, etc. The network model represented by Faster R-CNN has achieved good recognition results in image object detection applications. In image and video applications, deep learning technology is widely used in the field of unmanned driving of cars, video surveillance, target detection and recognition of drones and other fields. Due to the partial occlusion of the target, the blur caused by too fast movement, and the large deformation of the target in the scene, it often leads to problems such as low detection accuracy or no detection at all. In animation video detection and recognition, due to the exaggeration of ...

AI Technical Summary

Problems solved by technology

[0011] In view of the fact that the current deep convolutional neural network target detection algorithm based on a single image cannot detect the target with high precision in the case of motion blur, image quality degradation, partial occlusion of the detection target, and large deformation of the detection target, the present invention Extend the existing single image target detection algorithm, by fusing the different information of multiple frames before and after the image, overcome the low or even The problem of undetectable, improving the accuracy of object detection by utilizing multiple images

Images