Object classification method based on target characteristic graph

Abstract

The invention relates to an object classification method based on a target characteristic graph. The object classification method comprises the steps of respectively acquiring object multispectral data and spatial structure data in a to-be-detected area by means of a laser radar, a visible light image camera and a near-infrared image camera; then extracting to-be-detected ROI areas; then performing characteristic extraction on each ROI object area, and combining the characteristics for obtaining a characteristic word; and finally determining the characteristic by means of a deep learning classifier based on CNN, thereby realizing quick and reliable classification of the object. The data which are acquired by the multiple sensors are complementary to each other, thereby effectively preventing a problem of insufficient detecting capability of a single sensor. Multilayer spatial characteristic extraction realizes more comprehensive description to characteristics of the object. Furthermore higher reliability of a classification result of the deep learning classifier based on the CNN is realized.

Description

technical field [0001] The invention belongs to the field of intelligent video image processing, and in particular relates to an object classification method. Background technique [0002] Object automatic classification technology is widely used in agricultural production, industrial automation, resource recovery and other fields. The object classification method based on machine vision has the advantages of convenient installation, strong adaptability, non-destructive detection, etc., and is a current research hotspot. When there are many types and large numbers of objects to be classified in the scene, and there is a certain degree of mutual occlusion, how to robustly extract each object region and achieve accurate classification is a challenging task. [0003] In response to this problem, the literature [Application of MB-LBP features in visual target detection and classification, Master Thesis of Institute of Automation, Chinese Academy of Sciences, 2008] proposes to u...

AI Technical Summary

Problems solved by technology

[0003] In response to this problem, the literature [Application of MB-LBP features in visual target detection and classification, Master Thesis of Institute of Automation, Chinese Academy of Sciences, 2008] proposes to use MB-LBP features and EC0C rules to design multi-category target classification algorithms. This method is only applicable to It is used for the classification of large-sized moving objects such as vehicles and human bodies; the literature [Design and Implementation of Fine-grained Object Classification Method, Master Thesis of Beijing Jiaotong University, 2014] proposed an object classification algorithm using convolutional neural network, which realized the classification of horses, cattle, etc. Classification of larger animals such as sheep; patent [object classification method based on improved MFA and transfer learning for small sample sets, CN201510801292.3] discloses a small sample based on improved MFA (Marginal Fisher Analysis) and transfer learning Set (target domain) classification algorithm, which can recognize limited types of objects at the same time; patent [a smart object classification device with recognition function, CN201610089449.9] uses millimeter-wave radar to detect the shape of objects, and the objects to be classified need to be placed on a custom turntable Rotate up and down to achieve 360-degree shape acquisition, which greatly limits its application scenarios; the patent [A method for object classification based on video images, CN201510012901.7] discloses a method for object classification based on video images, including image segmentation, main Component analysis and rectangular saturation feature classification, this method can only better classify the human and vehicle targets in the video image; the patent [a target classification method and system based on the visual bag of words model, CN201410087579.X] by analyzing the sample pictures The position information and description information of feature points realize image classification, which is limited to the classification of specific two-dimensional plane objects

Images