A lightweight target detection method based on improved YOLOv3

By pruning and structural improvements to the YOLOv3 network model, including adjusting the residual structure and building branches, and using CIoU loss function and DIoU-NMS, the problem of difficult deployment of deep learning object detection algorithms in embedded systems is solved, and efficient and lightweight object detection is achieved, suitable for equipment with limited resources.

CN114707639BActive Publication Date: 2025-03-28CHONGQING UNIV
0 Cites 0 Cited by

Patent Information

Application Number
CN202210287339.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-03-22
Publication Date
2025-03-28
Estimated Expiration
2042-03-22

AI Technical Summary

Technical Problem

Deep learning object detection algorithms are difficult to deploy in embedded systems, mainly due to the huge amount of model operations and the difficulty in running on devices with limited computing resources and storage resources, which makes it difficult to meet real-time requirements in actual application scenarios.

Method used

By pruning the backbone part of the YOLOv3 network model, the original residual structure is adjusted to a sparse residual structure, and a convolutional branch and pooled branch are constructed, the loss function is adjusted to a CIoU loss function, and DIoU-NMS is used in the post-processing module to reduce the model size and maintain detection accuracy.

Benefits of technology

It achieves the reduction of the model size to 3.17MB while ensuring detection accuracy and real-time performance, which is 98.7% lower than the YOLOv3 algorithm. It has high lightweight and practicality, and is suitable for the deployment of embedded devices.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The present invention relates to a lightweight target detection method based on improved YOLOv3, the method comprising S1 collecting image data containing a target object to be detected, making a training set according to the VOC2012 data set format, S2 improving the backbone of an original YOLOv3 network model, S3 training the improved YOLOv3 network model, S4 inputting the image to be detected into the trained final improved YOLOv3 network model, and outputting the target object to be marked and displayed with an external rectangular frame. The present invention performs lightweight improvement on a traditional YOLOv3 algorithm, reduces the model size to 3.17MB while ensuring its accuracy and real-time performance, is simple to operate, easy to implement, and has high practicality.
Need to check novelty before this filing date? Find Prior Art