Method for detecting small target based on view angle conversion in sea-sky-line scene

Abstract

The invention discloses a method for detecting a small target based on view angle conversion in a sea-sky-line scene. The method comprises the following steps: acquiring a to-be-detected image; identifying a sea-sky-line; selecting an effective rectangular area of the to-be-detected image; transversely segmenting the effective rectangular region into N2 image blocks, wherein an overlapping region exists between two adjacent image blocks; arranging the N2 image blocks in N rows and N columns to obtain a recombined image; detecting sea surface targets in the recombined image by using a deep learning network model, obtaining first coordinate information of each sea surface target, and combining the first coordinate information of each sea surface target into a first coordinate information set; and converting the first coordinate information of each sea surface target into second coordinate information of the sea surface target in the to-be-detected image, and forming a second coordinate information set. According to the invention, the detection precision of small target ships and small target buoys near the sea-sky-line is improved through the selection of the effective area and the recombination of the image blocks.

Description

technical field [0001] The invention belongs to the technical field of target recognition, and in particular relates to a method for detecting small targets based on perspective conversion in a sea-planet scene. Background technique [0002] At present, target detection technology plays an increasingly important role in various fields, and it is also gradually mature. Most of the existing mature object detection techniques are based on deep learning methods. Usually, the deep learning method will scale the input image data to a fixed size in the image preprocessing stage, and then load the model to detect the target in the scaled image data, such as the widely used YOLO algorithm model and SSD algorithm model . [0003] In the field of sea surface target detection, small target ships and small target buoys near the sea antenna have extremely small targets due to their long distances. At present, when detecting small target ships and small target buoys near the sea line, t...

AI Technical Summary

Problems solved by technology

[0004] 1. In the long-distance sea-antenna image input to the deep learning network model, the long-distance sea-antenna image is a relatively high-resolution image, such as 5472*3648, but the pixels of small targets near the sea-antenna account for a relatively small proportion of the overall image. Most of the pixels are areas such as ocean waves and sky that users do not care about, so when running traditional algorithms, most of the computing time is applied to irrelevant areas; 2. In the image preprocessing stage of the deep learning network model, usually high The high-resolution image data is compressed to a resolution of 640*640 pixels, and even compressed to a resolution of 416*416 pixels for computational efficiency. The small targets near the sea antenna become smaller, and more interference pixels are introduced at the same time. As a result, the detection fails; 3. In the complex sea-antenna scene, the aircraft and birds appearing in the sky near the sea-antenna cause great interference to the identification of small target ships and small target buoys near the sea-antenna, and the interference cannot be detected. Excluded situations

Images