Improved YOLOv5-based image detection of cotton impurities

Abstract

Aiming at the embedded devices with limited resources and the characteristics of small and scattered distribution of cotton impurities, which cause the problems of low degree of accuracy and slow speed during detection, we propose the GBM-YOLOv5 lightweight cotton impurities detection model. First, the open-source software LabelImg was used to classify and label the targets in order to construct a dataset of cotton mottled images. Second, because of the lightweight structure of the Ghost module, the original CSP structure was replaced by the GhostBottleneck composed of the Ghost module, which effectively reduced the number of parameters and computation of the model. After the Ghost module, finally, a SoftPool structure was introduced in the SPP module to improve the pooling operation and retain more detailed feature information. Final results validate that the average detection accuracy of GBM-YOLOv5 was improved by 2.59% compared to YOLOv5 and 5.79% compared to YOLOv4 by training different network models on the well-constructed dataset. The model can meet the practical requirements of cotton industrial production and better improve the purity of cotton.