Target detection method based on selectable expansion convolution kernel size

Abstract

The invention discloses a target detection method based on a selectable expansion convolution kernel size, and relates to the field of computer vision and artificial intelligence. The method includes: firstly, extracting features through a convolutional neural network, and performing feature fusion through a feature pyramid; then, introducing a selectable expansion coefficient convolution module into a corresponding layer of the feature spectrum obtained through fusion, and obtaining better features through the module; and finally, carrying out multi-classification and bounding box regression on the fused feature layer, and continuously training an iteration model to obtain a target detection result after multi-scale fusion. According to the method, the precision is effectively improved, the real-time performance can be kept under an input picture with a certain size, and the method can be applied to places such as machine vision, face recognition, automatic driving, intelligent video and medical detection.

Description

technical field [0001] The invention relates to a target detection method based on a selectable dilated convolution kernel size, which belongs to the technical field of computer vision and intelligent information. Background technique [0002] Object detection is fundamental to computer vision tasks and to many applications in artificial intelligence. For target detection, its definition is as follows: Given an input RGB image, target detection completes two tasks: detection and recognition, that is, knowing what category the object belongs to and finding out where the object is located in the picture. Wherein, the category may be common species in nature, such as human, poultry, vehicle, etc., and a bounding box is used for positioning. Object detection has a wide range of applications in face recognition, automatic driving, human-computer interaction, content-based image retrieval, intelligent video surveillance, etc. [0003] Existing detectors are mainly divided into t...

AI Technical Summary

Problems solved by technology

The two-stage method divides the detection problem into two processes. First, region proposals are generated, and then candidate regions are classified and bounding box regression is performed. The beginning of this type of algorithm is the R-CNN algorithm proposed in 2016, but due to the two-stage A large number of borders will be generated, which greatly increases the computational complexity, so it is difficult to achieve real-time; while the single-stage detector adopts the idea based on regression, abandons the region proposal stage, and predicts the category probability and position coordinates of the object through the anchor point (Anchor) , the final detection result can be obtained through end-to-end learning, in which the abandonment of the region proposal stage can greatly reduce the computational complexity, so it can achieve real-time when the input resolution is appropriate. The representative algorithms are: YOLOv3, DSSD, RfineDet wait

Images