Lightweight anchor-frame-free target detection method for computer vision application

Abstract

The invention discloses a lightweight anchor-frame-free target detection method for computer vision application. The method comprises the following steps: extracting features of an input picture through a lightweight backbone network; constructing a lightweight feature pyramid structure by using a part of feature maps in the lightweight backbone network; mapping each pixel point coordinate of eachlayer of feature map on the feature pyramid back to the original map to obtain center point coordinates of all prediction frames; connecting a lightweight prediction branch behind each layer of feature map of the feature pyramid to obtain prediction frame information; decoding and calculating the prediction frame information obtained by all the prediction branches of the picture through a networkto obtain all prediction frames; and carrying out non-maximum suppression to obtain a final detection and identification result. According to the method, an anchor-frame-free design idea is adopted,all artificial design hyper-parameters and complex calculation caused by an anchor frame are eliminated, so that memory occupation during training is reduced, and the sensitivity of the model to targets of all scales is improved in combination with a feature pyramid structure, so the detection precision is improved.

Description

technical field [0001] The invention relates to the technical field of target detection, in particular to a lightweight anchor-free target detection method for computer vision applications. Background technique [0002] Target detection has always been an important problem in computer vision. Its main task is to automatically complete the prediction of the position and category of the target of interest in a picture through calculation. [0003] In computer vision application scenarios such as autonomous driving and drones, model lightweight is crucial for object detection algorithms. Therefore, in order to break through the limitations of the storage space and power consumption of the neural network model, the lightweight work of the model has been advancing. It is mainly divided into two methods: designing a lightweight neural network model and model compression, and the former has a higher priority. In recent years, computer vision researchers have proposed many cleverly...

AI Technical Summary

Problems solved by technology

[0004] However, these algorithms are all based on anchor boxes, and their main function is to explicitly enumerate a priori boxes of different scales and aspect ratios to predict information of different scales, which brings a lot of inconvenience: such as many problems introduced by anchor boxes. All hyperparameters require careful manual design, and the unbalanced ratio of positive and negative samples brought by a large number of anchor boxes, etc.

Images