Dense small target detection model construction method, model and detection method

Abstract

The invention provides a method for constructing a dense small target detection model, the model and the detection method. Based on the information fusion of the target midpoint context, through cutting the picture with high resolution, the method avoids the picture sampled under the input network from losing too much image information, and affects the network feature extraction. The residual pyramid feature extraction network is used to fuse the features of different scales, which improves the detection accuracy of the network for different size targets, especially for small targets. A RoIAlign layer instead of RoIPooling layer is used to solve the position deviation of candidate frames caused by feature mismatch of candidate regions. Because the small target features are easily lost in the network transmission, the center point context features are fused with the original RoI features, so that the network can make full use of the target context information, ensure the network runningspeed, more accurately locate and identify the dense small target, and improve the network performance.

Description

technical field [0001] The invention relates to a method for constructing a dense small target detection model, a model and a detection method, and relates to the field of target detection. Background technique [0002] Target detection is one of the most challenging topics in the field of computer vision at present. Its main task is to simultaneously recognize and locate the corresponding target in the picture according to a given picture or video. In recent years, with the rapid development of deep learning, object detection technology based on convolutional neural network has made significant progress, and has been widely used in autonomous driving, national defense security, medical field, human-computer interaction, etc. It is of great significance to the progress of human science and technology civilization. [0003] Among the traditional object detection methods, the deformable part model DPM (Deformable Part Model) is the most classic method of manually designing fe...

AI Technical Summary

Problems solved by technology

[0006] Although these methods have achieved good results, for situations where the targets in the picture are small, dense and occluded, the general target detection network framework will cause the feature information of small targets to be lost in the network transmission, and there are many false positive samples. Serious problems such as missed detection and wrong detection of small targets

Images