Depth learning target detection method based on pre-segmentation and regression

Abstract

The invention discloses a depth learning target detection method based on pre-segmentation and regression, mainly to solve the problems that the existing target detection method is poor in small target detection precision and long in detection time. The method comprises steps: 1) a quadtree segmentation algorithm is used to extract a region of interest in a to-be-detected image; 2) a basic convolution layer and an auxiliary convolution layer are used to carry out feature extraction on the region of interest, and feature graphs of multiple scales are obtained; 3) the position information of a default border is calculated on the feature graphs of multiple scales, a convolution filter is used for detection on the feature graphs of multiple scales, and multiple predicted borders and multiple category scores are obtained; and 4) non-maximum suppression is used for the multiple predicted borders and the multiple category scores, and the final target border position and the category information are obtained. A small target in the image can be quickly and accurately detected, and the method can be used for target real-time detection in unmanned aerial vehicle aerial photographing.

Description

technical field [0001] The invention belongs to the field of image information processing, and specifically relates to a deep learning target detection method, which can be used for accurate real-time positioning and classification of targets. Background technique [0002] Target detection is a challenging topic in the field of computer vision. Its core task is to use some target recognition algorithm and search strategy in static pictures or videos to obtain the position and category of specific targets in the image or video. At present, the methods of target detection are mainly divided into target detection algorithms based on features and machine learning and detection methods based on deep learning. Among them, the method based on feature and machine learning is to realize target detection through the process of region selection, feature extraction, classifier classification and other processes. Region selection is to traverse the entire image through sliding windows t...

AI Technical Summary

Problems solved by technology

SPP-NET, Fast R-CNN and Faster R-CNN are produced by R-CNN gradually optimizing and speeding up. The accuracy and speed of target detection have been greatly improved. steps and the positioning time is too long, so the target detection cannot be performed in real time

The representative deep learning target detection algorithms based on the regression method are YOLO and SSD. This type of algorithm mainly predicts the position and category of the target directly from the image to be detected through the regression method. This method greatly speeds up the target detection speed and can reach Requirements for real-time target detection, but there are strict requirements on the size of the input image and the target position positioning is poor, and it is impossible to detect small targets in the image

YOLO and SSD300 require the input image size to be 448*448 and 300*300 respectively. If the image to be detected is reduced to a specific size, image details will be lost, resulting in the inability to detect small targets.

Images