Small target detection method for detecting aerial photo

Abstract

The invention relates to a small target detection method for detecting an aerial picture. The method comprises the following steps of: 1, establishing a data set; 2, designing an algorithm; 3, performing iterative training on the data until the loss function value is minimum; 4, performing model parameter adjustment and fine adjustment on the detection model, and optimizing part of hyper-parameters; and 5, verifying the training model on a small target data set, comparing with a design index, and selecting wrong data or properly adjusting parameters. The invention aims to provide a lightweight small target detection method combining the context and the attention mechanism, which takes different feature information as the context and focuses a target in an image so as to effectively improve the detection speed and remarkably compress the calculation time.

Description

technical field [0001] The invention belongs to the field of computer vision, in particular to a small target detection method for detecting aerial pictures. Background technique [0002] In the aerial images, the target pixels are small, the resolution is low, the proportion of the overall area of ​​the image is small, and the background is complex and diverse, which brings great troubles to the detection and positioning. [0003] At present, there are usually the following research ideas for small target detection problems: [0004] Data enhancement: Through different enhancement strategies such as image rotation, translation, cropping, adaptive sampling, and adding noise, increasing the amount of data and enriching image types will help improve the robustness and generalization ability of the training model. The detection ability of the neural network is largely attributed to the scale and quality of the data set, so data enhancement is the most direct and easy way to im...

AI Technical Summary

Problems solved by technology

According to the solution of context and attention mechanism, the effectiveness of the method only exists when the context-related information is tight, and objects that are easy to detect do not always really help the discovery of small objects. The encoder-decoder framework in the attention mechanism needs Compress the entire sequence information into a fixed-length vector, causing the decoder to fail to acquire enough information in a timely and accurate manner

[0008] The patent application publication number CN112507840A discloses a human-computer hybrid enhanced small target detection and tracking method and system; it relies on the "eye movement" attention device of wearable devices, and needs to find the pupil center and corneal reflection center as positioning Although the basic point improves the accuracy of identification and tracking, it is only suitable for some wearable devices and has a limited scope of application

Images