Image detection method and device, equipment and storage medium

Abstract

The invention provides an image detection method and device, equipment and a storage medium, and relates to the technical field of detection and various application scenes such as cloud technology, artificial intelligence, intelligent traffic and auxiliary driving. The method comprises the following steps: extracting features of a target image so as to obtain N feature maps with different scales, wherein N is a positive integer larger than 1; determining N weight coefficients respectively corresponding to the N feature maps, wherein the N weight coefficients are respectively used for representing importance degrees of the N feature maps in a detection process of a target object in the target image; based on the N weight coefficients, determining M feature maps in the N feature maps, wherein M is a positive integer smaller than or equal to N; and detecting the target object based on the M feature maps. According to the method, on the premise that an accuracy rate and a recall rate are guaranteed, calculation overhead is reduced, the requirement of algorithm deployment for the equipment computing power is lowered, and lightweight deployment of the equipment is achieved.

Description

technical field [0001] The embodiments of the present application relate to the technical field of detection, and more specifically, to an image detection method, device, device, and storage medium. Background technique [0002] So far, the existing target detection training frameworks all adopt the method of feature pyramid (feature maps of different feature scales), and fuse the feature maps of different downsampling scales in a certain way to obtain the fused feature map, and then fuse The final feature maps are all input to the subsequent classification network or regression network to obtain the position of the target and give the category of the target; in short, the current target detection scheme directly uses the feature maps of each different downsampling scale as the subsequent recognition target The class of the object and the input to identify the location of the target object. [0003] However, generally speaking, the computing power of mobile devices is often...

AI Technical Summary

Problems solved by technology

[0003] However, generally speaking, the computing power of mobile devices is often much smaller than that provided by the cloud. Therefore, the shortcomings of existing target detection solutions are mainly reflected in mobile deployment, because the algorithm directly uses all different downsampling scale feature map, so deploying it on a mobile device will have a high computational overhead

Images