Small-scale target detection method based on weak edge

Abstract

The invention discloses a small-scale target detection method based on a weak edge, and the method comprises the following steps: 1), establishing a weak edge detection network: enabling a backbone network to be divided into five stages based on VGG16, and obtaining the features of different scales; fusing the results of each stage to obtain a final result of edge detection; 2) constructing loss functions of the trunk network and the bypass branches, and performing weighted summation on the auxiliary loss functions of the plurality of branches and the loss functions on the fusion features; 3)connected region analysis: carrying out binarization processing on the generated edge images; determining a minimum bounding box, and generating a multi-instance suggestion area; and 4) detecting a small-size target by using the RPN, summarizing the detection results of all multi-instance suggested regions, and further adjusting the target detection result by using the Fast R-CNN. According to themethod, a good detection result is obtained when the situation that the target is small in the image and the situation that the targets are mutually adhered and shielded are processed.

Description

technical field [0001] The invention relates to a target detection task in computer vision, and is a small-sized target detection framework capable of performing detection and positioning functions on objects with small scales in an image, belonging to the field of image processing. Background technique [0002] Object detection is a prerequisite for numerous tasks in computer vision, such as face recognition, pedestrian detection, semantic segmentation, object tracking, etc. Object detection is to judge whether there is an object of interest in the picture, and then express it through other forms such as detection frame. According to the size of the object in the picture, the target detection task can be divided into conventional target detection and small-size target detection. Small targets refer to targets with a small ratio of their own size to the size of the image. Small target detection is a difficult point and development trend in target detection. It has many appl...

AI Technical Summary

Problems solved by technology

Zhu et al. used the edges to generate suggested regions, and then classified them region by region, which effectively improved the detection accuracy of traffic signs. However, the background of the data set applied by this method is simple, and it cannot handle the small target detection in the case of adhesion and occlusion.

Images