Recyclable solid waste detection based on image fusion and convolutional neural network

The most solid waste image datasets usually contain only a single object with a plain background, which is quite different from the real environment. In addition, the waste images labeling process takes a long time and is labor cost. To address these problems, we proposed an effective method to extend the dataset based on image fusion. Herein, we use image fusion technology to make a recyclable solid waste dataset Trash-Fusion automatically, where the images contain different categories of objects with complex background, and all classification and location labels are collected in the process of image fusion. Moreover, an actual scene dataset Trash-Collect is constructed, images of which are downloaded from the Internet or collected by ourselves. A mixed dataset of Trash-Fusion and Trash-Collect is sent to several convolutional neural networks for training, and YOLO v5 achieves the highest detection precision with 60 FPS.