New Benchmark for Household Garbage Image Recognition

Household garbage images are usually faced with complex backgrounds, variable illuminations, diverse angles, and changeable shapes, which bring a great difficulty in image classification. Due to the ability discover problem-specific features, deep learning especially convolutional neural networks (CNNs) have been successfully widely used for representation learning. However, available stable household datasets insufficient, seriously limits development of research application. Besides, state-of-the-art field classification is not entirely clear. To solve this problem, study, we built new open benchmark dataset by simulating different lightings, shapes. This named 30 classes (HGI-30), contains 18 000 classes. The publicly HGI-30 allows researchers develop accurate robust methods recognition. We also conducted experiments performance analyses CNN on HGI-30, serves as baseline results benchmark.