Crop pest recognition in natural scenes using convolutional neural networks

•A manually collected and validated crop pests dataset (10 classes/5629images).•A fine-tuned GoogLeNet model for the identification of 10 crop pests types.•Combine several pre-processing techniques to deal with the complex background.•The average detection accuracy for the proposed model is over 98%.•The fine-tuned GoogLeNet model obtained an improvement of 6.22% compared to the state-of-the-art method. Crop diseases and insect pests are major agricultural problems worldwide, because the severity and extent of their occurrence causes significant crop losses. In addition, traditional crop pests recognition methods are limited, ineffective, and time-consuming due to the manual selection of the useful feature sets. This paper introduces a crop pest recognition method that accurately recognizes ten common species of crop pests by applying several deep convolutional neural networks (CNNs). The main contributions of this paper are (1) a manually collected and validated crop pest dataset is described and shared; (2) a fine-tuned GoogLeNet model is proposed to deal with the complicated backgrounds presented by farmland scenes, with pest classification results better than the original model; and (3) the fine-tuned GoogLeNet model obtains an improvement of 6.22% compared to the state-of-the-art method. As a result, the proposed model has the potential to be applied in real-world applications and further motivate research on crop disease identification.