Waste Image Segmentation Using Convolutional Neural Network Encoder-Decoder with SegNet Architecture

Sustainable waste management has become one of the great concerns in many countries to reduce the negative environmental and health impact caused by the increase of unmanaged household garbage. One of the key elements is waste recycling, that relies on the process of garbage sorting to separate the recycled garbage into different categories for further processing. The sorting process in waste recycling is usually done manually by hand-picking. Therefore, a system that can recognize waste automatically is needed so that the waste sorting process can be done more quickly and accurately. In this paper, we propose a waste segmentation method using Convolutional Neural Network based on the Encoder-Decoder approach of SegNet architecture [5]. We compare two different setups of the architecture based on the number of filters in each convolutional layer, then evaluate our model using TrashNet benchmark dataset. Our experiment shows that one of our proposed architecture managed to achieve 82.95% intersection over union (IoU) value, which is higher than the previous work by the TrashNet developer.