Original scientific paper

https://doi.org/10.17559/TV-20220420124810

An Improved ResNet-50 for Garbage Image Classification

Xiaoxuan Ma orcid.org/0000-0002-8221-4031 ; Beijing University of Civil Engineering and Architecture, School of Electrical and Information Engineering, No. 15, Yongyuan Road, Huangcun, Daxing District, Beijing, China
Zhiwen Li orcid.org/0000-0001-6557-0774 ; Beijing University of Civil Engineering and Architecture, School of Electrical and Information Engineering, No. 15, Yongyuan Road, Huangcun, Daxing District, Beijing, China
Lei Zhang orcid.org/0000-0002-8221-4031 ; Beijing University of Civil Engineering and Architecture, School of Electrical and Information Engineering, No. 15, Yongyuan Road, Huangcun, Daxing District, Beijing, China

Abstract

In order to solve the classification model's shortcomings, this study suggests a new trash classification model that is generated by altering the structure of the ResNet-50 network. The improvement is divided into two sections. The first section is to change the residual block. To filter the input features, the attention module is inserted into the residual block. Simultaneously, the downsampling process in the residual block is changed to decrease information loss. The second section is multi-scale feature fusion. To optimize feature usage, horizontal and vertical multi-scale feature fusion is integrated to the primary network structure. Because of the filtering and reuse of image features, the enhanced model can achieve higher classification performance than existing models for small data sets with few samples. The experimental results show that the modified model outperforms the original ResNet-50 model on the TrashNet dataset by 7.62% and is more robust. In the meanwhile, our model is more accurate than other advanced methods.

Keywords

attention module; garbage classification; multi-scale feature fusion; ResNet

Hrčak ID:

281668

URI

https://hrcak.srce.hr/281668

Publication date:

15.10.2022.

Visits: 4.183 *