Image based object recognizing method

Abstract

The invention provides an image based object recognizing method. The method includes a training process used for establishing a first database including first characteristic vectors used for describing object shapes and a second database including second characteristic vectors used for describing object types; a recognition process. The recognition process includes inputting pictures into a depth convolutional neural network; generating at least one candidate box in each picture, performing pooling treatment on a characteristic map corresponding to each candidate box so as to obtaining third characteristic vectors; comparing the third characteristic vectors with the first characteristic vectors in the first database, calculating coefficient of association between the two vectors, and selecting the candidate boxes corresponding to the third characteristic vectors to be bounding boxess when the coefficient of association is greater than or equal to a specific threshold value; inputting the images in the bounding boxes to a classification neural network so as to obtain fourth characteristic vectors; and based on the fourth characteristic vectors, the second characteristic vectors and the second database, performing a kNN (k-Nearest Neighbor) sorting algorithm for recognizing object types.

Description

technical field [0001] The invention relates to an image-based object recognition method. Background technique [0002] In recent years, deep convolutional neural networks have made significant progress in the fields of object recognition, object localization, and image segmentation. Through the object recognition algorithm based on the deep convolutional neural network, the recognition accuracy of the machine has even surpassed that of humans in individual tasks. [0003] Other algorithms, such as the R-CNN algorithm, Faster R-CNN algorithm, YOLO (you only look once) algorithm, SSD algorithm and R-FCN algorithm disclosed in the prior art, have also achieved great success in the fields of object positioning and image segmentation. , a higher accuracy was obtained. [0004] However, these methods generally have fewer types of recognized objects (20-80 types), and for new object types, a large number of labeled training sets are required, and it takes a lot of time to retrai...

AI Technical Summary

Problems solved by technology

[0004] However, these methods generally recognize fewer types of objects (20-80 types), and for new object types, a large number of labeled training sets are required, and it takes a lot of time to retrain the neural network to achieve the recognition effect.

In addition, the vast majority of object positioning and recognition algorithms lack the ability to distinguish between different individuals of the same kind of objects

Images