Small target detection method and system under low visibility

Abstract

The invention provides a small target detection method and system under low visibility. The method includes: taking RefineDet as a detector, and processing the target image by adopting multi-scale retinal enhancement with color restoration to obtain an enhanced image; filtering a negative anchor point in the enhanced image by using the anchor point refining model, establishing an association relationship between the anchor point refining model and the target detection model, sharing a feature image in the anchor point refining model to the target detection model based on the association relationship for regression and category prediction, and outputting an identification result. Under the conditions of low brightness, backlight and fuzzy environment, a multi-scale retina enhancement algorithm with color recovery is adopted to reduce the sensitivity of the image under different illumination, reduce noise and improve the characteristics of the small target after being extracted by the convolutional neural network; through fusion of the high-level semantic features and the low-level features, the features with rich receptive fields, sufficient details and abstract content are obtained, and the discrimination degree of small targets is improved.

Description

technical field [0001] The present invention relates to the technical field of computer vision, in particular to a small target detection method and system in low visibility. Background technique [0002] The small target problem has always been a difficulty in visual tasks such as object detection and semantic segmentation. The detection accuracy of small targets is usually only half of that of large targets. In the MS COCO dataset, a small target is defined as an object with an area smaller than 32*32. Due to the low resolution, blurred images and less information carried by small targets, the feature expression ability is weak, and it is difficult to extract more effective features, making it difficult to detect small targets. At present, the more popular methods to solve the problem of small target detection are the scale enlargement method, which is to enlarge the image before detection; the Anchor window improvement method, which is to improve the density, range and s...

AI Technical Summary

Problems solved by technology

[0004] Domestic patent CN110097129A discloses a remote sensing target detection method based on contourlet grouping feature pyramid convolution, which solves the problems of low recall rate and high false alarm rate when the edges of optical remote sensing images of aircraft and ships are blurred

Images