Research On Insect Image Recognition Based On Deep Learning

Insect image recognition is a branch of insect recognition field and it is one of the key technologies for biological research,pest monitoring and detection.Insect image recognition relies on the extracted and abstracted features of objects in images.The traditional methods of image-based insect recognition always have complex preprocessing processes.It needs to design a corresponding feature extraction algorithm according to the specific task requirements,and the accuracies are relatively low.Deep learning methods,which provide ways to approach the appropriate and high-complexity feature extraction algorithms in the specific tasks,are used to improve the performances of models in recognizing the images of insects in this thesis.The researches conducted in a combination of theoretical analysis and experimental comparison in this thesis are carried out in the order of collecting samples,selecting models,improving models and training algorithms,classification,and object detection.The main works and results in this thesis are shown as follows:(1)An image sample set containing 10 species of Cerambycidae is collected,and a total of 2,782 natural background insect images are annotated.The sample set is extended by spatial transformation algorithms,and the expanded sample size reached 12,474.The sample set is versatile and can be used in the researches of related technologies for insect identification and image recognition.(2)Based on the insect image sample set,Dense Net is selected as the basic network for the study of insect image classification in this thesis through theoretical analysis and experimental comparison.And then,an improved scheme named UDenseNet is designed.The improved model enhances feature transfer and reuse,and uses a smaller growth rate to reduce network parameters and alleviate over-fitting of the network model on the insect image sample set.About 2/3 convolution parameters are removed in the improved model and the accuracy of classification improves about 0.5%.Furthermore,a mixed pooling algorithm in which the features of max pooling and mean pooling are summed by weights is designed to enhance the extraction of useful features through parameters learning.The accuracies of DenseNet and UDenseNet with mixed pooling structures are improved on the insect image sample set.(3)An adaptive learning rate descent algorithm which is based on the Stochastic Gradient Descent algorithm is proposed.This algorithm achieves the goal of automatically decreasing the learning rate by detecting the bottleneck of training.The experiments in this thesis show that training with this algorithm can avoid the processes of repeatedly adjusting the step size of learning rate descent in specific tasks,and obtain relatively good generalization.(4)The fine-tuning of DenseNet and direct training of DenseNet and UDenseNet are performed using the improved training algorithm.And then the trained models are used to classify insect images.The recognition results are enhanced by using multi-crop strategy,and the fine-tuned model of DenseNet achieves an accuracy of 99.83% on the extended test dataset which contains 600 insect images.In addition,in order to solve the situation that an image contains multiple types of detection targets in actual applications and complete the insect image recognition solution,the target detection framework R-FCN is used to perform model training and object detecting on the manually annotated sample set of insect images.The model trained by the improved training algorithm achieves a mAP of 90.51% on the validation dataset.