Improved ShuffleNetV2 for Garbage Image Classification

This paper proposes a novel lightweight classification network based on ShuffleNet V2 to promote the accuracy of garbage image classification. Compared with the conventional network, there is a package of optimizations to both ameliorate the classification accuracy and inference speed of the model. Firstly, the optimized network introduces the Leaky ReLU as a substitute for the activation function. It also adopts the squeeze and excitation network to enhance important features. In addition, the flattened layer is applied to downscale the data, and three fully connected layers are used for enhancing the learning capability. Furthermore, the model is pre-trained based on transfer learning, wherein the feature knowledge learned from ImageNet datasets is transferred to garbage classification task. Extensive experiments on the garbage image dataset verify the overall performance, which improves the classification accuracy by 6.3% with a 35.4% reduction in time consumption.