Target detection method combining positioning information and classification information

Abstract

The invention provides a target detection method combining positioning information and classification information. On the basis of the Faster R-CNN, the classification information and the positioninginformation required by target detection are comprehensively considered, the detection precision can be improved to a great extent, and the detection speed is ensured. Firstly, in order to more fullyextract classification features, reduce parameters required by operation and reduce the possibility of network overfitting, a full-connection operation module of an inverse residual structure is provided; secondly, in order to enhance the positioning capability, an operation module combining positioning information and classification information is provided. The whole method adopts a back propagation algorithm for training and introduces an online difficult case analysis technology during training so as to further improve the precision. Experiments show that compared with the Faster R-CNN, themethod has great advantages in speed and precision, and has great application value.

Description

Technical field [0001] The invention relates to the technical field of image processing and artificial intelligence, and in particular to a two-stage target detection method based on a deep convolutional neural network. Background technique [0002] As an important part of the computer vision field, target detection has always received great attention from scientific researchers. In recent years, with the advancement of artificial intelligence, especially deep learning algorithms represented by deep convolutional neural networks, the field of target detection has also made huge developments-target detectors based on deep convolutional neural networks have basically replaced the traditional In the machine learning era, the method of target detection in the "Hog+SVM" mode. Compared with traditional machine learning algorithms, deep convolutional neural networks can autonomously extract more essential and representative features of multiple targets, and thus have better accuracy an...

AI Technical Summary

Problems solved by technology

Although this change can improve the accuracy to a certain extent, the detection speed is greatly reduced.

On the basis of Faster R-CNN, in order to further improve the accuracy, many researchers have carried out the work of multi-scale feature combination: by combining the shallow network part containing more edge information and the deep network part containing more classification information to carry out target Detection such as Hyper-Net, FPN, etc., but these detectors require a lot of GPU memory and are very slow at runtime

Images