Unsupervised Cross-Domain Leaf Counting With Minimizing Feature Distances

Herbs are very important for human beings.They can be used as medicine or as food to supplement the daily vitamin and calcium needed by the body.Monitoring the growth of plant leaves can effectively predict the final yield.As a still used leaf counting method,the human eye measurement method often leads to poor counting results due to human eye fatigue.In addition,traditional supervised tasks have the weakness of poor generalization ability,and require a large number of training datasets to solve tasks in a specific scenario.In this paper,we propose a method that can deal with leaf counting tasks in different scenarios without relabeling specific datasets and saving a lot of labor costs.This method is based on the density map estimation method.Only using the public Arabidopsis leaf dataset with label,the cross domain leaf counting task can be completed on the other two Arabidopsis and Isatis indigotica datasets without label.In order to solve the problem of low accuracy of existing unsupervised cross domain counting methods,we proposes a method that can effectively reduce the feature difference between the source domain and target domain,which to some extent solves the accuracy problem of cross domain leaf counting and the defect of feature mapping.The main work of this paper is as follows:(1)Aiming at the poor generalization ability of the traditional supervised leaf counting method,a new unsupervised cross domain solution is proposed.This method has the following characteristics: a)For the target domain,it is an unsupervised method,which does not need to manually label the images collected in the target domain;b)For the same task in multiple different scenarios,it has the generalization ability that ordinary supervised tasks cannot achieve;c)The shortcomings of feature mapping of existing unsupervised cross domain leaf counting methods are improved,and the counting accuracy is improved.(2)We improve the generator of Cycle GAN to generate "fake data" with high quality.At the same time,this paper adopts Bayesian loss function probability density map estimation method based on density map generation without Gaussian kernel function,which makes the model have better density map generation performance.(3)Tests were conducted on three open leaf datasets.The experimental qualitative and quantitative results show that the accuracy of this method has been greatly improved on the two target domain datasets,which is higher than the advanced cross domain leaf counting method on the same dataset.