Insect pests are
known to be a major cause of damage to important
agricultural crops. We present a deep learning-based
pipeline for the visual localization and counting of
agricultural pests by self-learning of a saliency feature
map. The method integrates one convolutional neural network
of ZF (Zeiler and Fergus model) and a region proposal
network with Non-Maximum Suppression (NMS) to remove
overlapping detections. First, the convolutional layers of
ZF Net, which removes the average pooling layer and fc
layers, were used to compute feature maps, which better
retained the original pixel information through smaller
convolution kernels. Then, several critical parameters were
optimized, including the output size, score threshold, NMS
threshold, and so on. To demonstrate the practical utility,
different feature extraction networks were explored,
including AlexNet, ResNet and ZF Net. Finally, the original
images were tested by the trained model with smaller
multi-scale images, and the model achieved a precision of
0.93 with a miss rate of 0.10. Additionally, our model
achieved a mean Accuracy Precision (mAP) of 0.885. |