Recognition and counting of typical apple pests based on deep learning

The recognition and counting of apple pests sampled by different sex attractants are very important and significant for pest control. Convolutional neural networks (CNNs) are common artificial intelligence algorithms widely used in image recognition and counting. However, because the sizes of different species of pests are different and the densities of pests on the sticky board are sometimes considerable, it is difficult to recognise and count them accurately and efficiently. This study proposes an improved recognition and counting approach based on deep learning with data reorganisation, termed 'MPest-RCNN'. The contributions herein are twofold: (1) A new structure of Faster R-CNN is proposed by using ResNet101 feature extractor which has higher precision of recognition. (2) We propose a new convolutional network structure with small anchors to extract features such that the recognition accuracy is improved for small pests. We took three typical pests in apple orchards to establish an original data set using sex attractants. The proposed MPest-RCNN model solves the recognition problem of multiple types and sizes of pests by using different sex attractants. Finally, experiments are conducted, and a comparative analysis is provided for the proposed approach. The experimental results demonstrate that the precision, sensitivity, specificity, and F1-Score of the proposed approach reach 99.11%, 99.88%, 99.42%, and 99.50% respectively. In contrast with Faster R-CNN, the precision, sensitivity, and F1-Score increase by 0.31%, 7.77%, and 4.25% respectively. The comparative experimental results demonstrate that the mean average precision (mAP) of the proposed approach is higher than that of the currently used pest recognition approaches. In addition to reducing the complexity of creating multiple recognition models for different kinds of pests, the proposed model shows promise as an effective means for recognising and monitoring of other targets with similar characteristics, thereby providing theoretical support for ecological informatics in the future.