Quick detection method for objective on the basis of multi-scale characteristic pattern

Abstract

The invention discloses a quick detection method for an objective on the basis of a multi-scale characteristic pattern. The method comprises the following steps that: firstly, through a convolutionalneural network, automatically extracting the multi-scale characteristic pattern to avoid a complex characteristic design and extraction process in a traditional method; secondly, putting forward an effective characteristic pattern fusion method by considering a situation that characteristic expressions learnt by different convolutional layers are different, realizing by a lightweight compression type bilinear function to improve characteristic pattern fusion efficiency and enrich context information, and on the basis, combining the multi-scale characteristic pattern with a channel attention mechanism to highlight useful information and inhibit redundant information so as to further enhance characteristic pattern representation ability; and finally, using the enhanced multi-scale characteristic pattern for objective detection, and obtaining an optimal model through multiple iterative training. Compared with the prior art, the method which is put forward by the invention lowers time costas far as possible while detection accuracy is improved, the quick detection of the objective is realized, and the method has a wide application prospect on the aspects of mobile robots, automatic driving, intelligent video surveillance and the like.

Description

technical field [0001] The invention relates to a fast target detection method based on a multi-scale feature map, which belongs to the field of computer vision and intelligent information processing. Background technique [0002] Object detection, as the basis of other high-level visual processing and analysis tasks, has always been one of the core issues in the field of computer vision. It has important economic and practical value and great potential in many aspects such as mobile robots, autonomous driving, and intelligent video surveillance systems. For massive image data and dynamically changing scenes, it is of great significance to efficiently acquire and identify various objects of interest and achieve accurate and fast object detection. [0003] Traditional object detection methods rely on artificially designed features, and input the extracted features into a classifier for classification and recognition. However, artificially designed features have many limitati...

AI Technical Summary

Problems solved by technology

The method without region proposal breaks through the bottleneck of real-time performance, but the detection effect is limited, and the detection accuracy of groups close to each other and small targets is not high

Images