Multi-scale detection of stored-grain insects for intelligent monitoring

In order to implement intelligent monitoring for insects in grain warehouses, a Multi-Scale Insect Detector (MSI_Detector) was developed by applying deep convolutional neural networks. It solved the problem existing in common anchor-based insect detection methods whose performance decrease sharply as insects become smaller. We built a feature pyramid network to extract insect image features with different spatial resolutions and semantic information, and tiled anchors with reasonable scales on each pyramid level to handle different scales of insects well. Besides, we altered the classification and box regression subnets to residual structure in order to improve the detection performance, and proposed the recombined loss function to balance the weights of easy and hard samples during training for both insect classification and box regression tasks. Excellent results for the detection of adults of 10 species 6 genera of common stored-grain insects were achieved, the mean Average Precision reached 94.77%, which demonstrated the robustness to insect scales.