Method and device for improving neural network target detection performance

Abstract

The invention discloses a method and device for improving the target detection performance of a detection neural network, and relates to the technical field of digital image processing. The method comprises the following steps: determining the sizes of a plurality of rectangular sliding windows for scanning according to the input size of a target detection neural network algorithm and the size ofan original input image; during detection of each frame, rotating the sliding window sub-image in each rectangular sliding window according to a preset angle and performing zooming processing to generate a sliding window rotation mapping sub-image, and performing zooming processing on the original input image to generate a full-image mapping sub-image; combining and splicing the full-image mappingsub-image and each sliding window rotation mapping sub-image to form a detection input image; and detecting the detection input image through a target detection neural network algorithm correspondingto the input scale. According to the invention, the computing power and bandwidth requirements of the target detection algorithm on the monitoring edge computing equipment are reduced, and the targetdetection distance and the detection accuracy are optimized.

Description

technical field [0001] The invention relates to the technical field of digital image processing, in particular to a method and device for improving the detection performance of neural network targets. Background technique [0002] With the rapid development of artificial intelligence and deep learning technology, in the monitoring field, the target detection method based on the convolutional neural network (CNN) algorithm has been widely used. The commonly used target detection process is to slide from left to right and from top to bottom window, using classification to identify objects. To detect different object types at different viewing distances, we can use windows of different sizes and aspect ratios (sliding windows). Among the target detection algorithms, commonly used ones such as RCNN, Fast RCNN and Faster RCNN are methods based on candidate regions and deep learning classification. From RCNN to Fast RCNN and then to Faster RCNN, mAP (mean Average Precision) is co...

AI Technical Summary

Problems solved by technology

However, in a few special monitoring occasions, there will be abnormal objects rotating at 90 degrees, 180 degrees or 270 degrees, such as horizontal (lying or lying) human bodies in the swimming pool scene, and inverted human bodies (human body) in the gym scene. Face), etc., the existing target detection algorithm has a high miss rate for the above targets

This is because most of the training targets are upright, which leads to the detection performance of the trained detection algorithm for horizontal (including lying and lying down) and inverted targets is not ideal.

In order to solve the above problems, the conventional method in the prior art is to rotate the monitoring input image by 90 degrees, 180 degrees and 270 degrees and then re-detect it, so as to improve the detection rate of these targets, but after adopting the above-mentioned scheme, it needs to be processed The amount of data will increase by more than 4 times compared with the original, which puts forward higher demands on the computing power and bandwidth of the equipment

[0006] To sum up, the existing low-complexity optimization methods cannot achieve comprehensive results in terms of detection accuracy, frame rate, farthest and shortest detection distances, etc.

However, the actual monitoring application scenarios are complex, and the monitoring equipment needs to meet high target detection accuracy, and also needs enough frame rate to detect fast passing targets, and also needs to detect large-scale targets in the vicinity and small-scale targets in the distance. can be detected (when the target is close to the camera or passes by in the distance), especially for the detection of abnormal targets

Images