RecycleTrashNet: Strengthening Training Efficiency for Trash Classification via Composite Pooling
Abstract
In this paper, we propose RecycleTrashNet to classify house trash based on deep neural networks. In our model, we use 3×3 filter in convolution layer instead of 7×7 filter, which is a smaller filter tends to learn more features. Since single max pooling or average pooling in pooling layer can’t achieve good performance, we present composite pooling to preserve as many image features as possible. Experiments on trash dataset from Stanford University demonstrate good performance of RecycleTrashNet over other neural network models in speed and accuracy. It can improve training efficiency, which achieves classification results with 88% test accuracy at only 80 epochs.
Keywords
Deep residual network, Composite pooling, Trash classification, Residual block
DOI
10.12783/dtetr/acaai2020/34190
10.12783/dtetr/acaai2020/34190
Refbacks
- There are currently no refbacks.