DeepWeeds: A Multiclass Weed Species Image Dataset for Deep Learning

by:BrightMart     2020-03-03
Recently, more and more research has been done on robotic weed control, which has the potential to increase agricultural productivity.
Most of the work focuses on the development of robots for farmland, ignoring the problem of weed management faced by farm livestock farmers.
Perhaps the biggest obstacle to widespread adoption of robotic weed control is the robust classification of weed species in the natural environment.
The unparalleled success of deep learning makes it an ideal candidate for identifying various weed species in complex pasture environments.
This work contributes to the first large, public multi-category weed image dataset from Australian pastures;
Robust classification methods are allowed to be developed to make robotic weed control feasible.
The DeepWeeds data set consists of 17,509 labeled images of important weed species in eight countries from eight locations in northern Australia.
This paper uses the benchmark deep learning model Inception-presents a baseline for classification performance on a datasetv3 and ResNet-50.
The average classification accuracy of these models reached 95. 1% and 95.
7% respectively.
We also show real-time performance of ResNet
The average reasoning time of 50 architecture is 53. 4 ms per image.
These strong results indicate a good prospect for implementing robotic weed control methods in Australian pastures in the future.
Robot weed control is expected to take a step
Changes in agricultural productivity.
The main benefit of the autonomous weed control system is the reduction of labor costs, while it is also possible to reduce the use of the herbicide by more effectively selectively applying it to weed targets.
Improving weed control will have a huge economic impact.
It is estimated that in Australia alone, farmers will spend an Australian dollar.
$5 billion a year is spent on weed control and $2 is lost.
Agricultural production affected accounted for 5 billion per cent.
The successful development of agricultural robot technology has the potential to reduce these losses and increase productivity.
Research on robotic weed control focuses on four core technologies that many consider: detection, mapping, guidance, and control.
Among them, detection and classification is still an important obstacle to the commercial development and industry acceptance of robot weed control technology.
There are three main detection methods that focus on different representation of the spectrum.
Various successes have been achieved in using imagesSpectrum based
Spectral Image Based
Weed Identification method based on ground and aerial photography.
Spectral and spectral images
Method-based best suited for highly controlled sites
Specific environments, such as arable farmland, where the spectrometer can be customized according to the environment for consistent collection and testing.
However, the harsh and complex pasture environment makes
Image-based approach is difficult to implement
Image-based methods can obtain cheaper and simpler images in different light conditions, especially when deployed on mobile vehicles in real time.
So for this work, we focus on images
Basic techniques for identifying weed species.
Using computer vision algorithm to realize automatic identification of plants is an important theoretical and practical challenge.
One way to address this challenge is to identify plants from leaf images.
To solve this problem, several algorithms and methods have been developed.
Perhaps the most promising leaf of the recent
Classification methods based on deep learning models, such as convolutional neural networks (CNN);
It is now dominant in many computer vision-related fields.
For example, the large-scale Visual Recognition Challenge of ImageNet (ILSVRC)
Since CNN\'s first win in 2012, it has been dominated by CNN variants and there is a big gap.
This and other recent successes should use deep learning in the detection and classification of weed species.
From linear regression to CNNs, the performance of each machine learning model is constrained by the data set it is learning.
There are many data sets of weed and plant life images in the literature.
The annual life cycle plant identification Challenge provides a set of 2015 data sets consisting of 113,205 images belonging to 41,794 observations of 1,000 trees, herbs and ferns.
This huge data set is very unique and most of the other works show the site
Specific data sets of weeds they are interested in.
These methods provide high classification accuracy for the target data set.
However, most data sets capture their target plant life in perfect laboratory conditions.
While perfect laboratory conditions allow for robust theoretical classification results, deploying a classification model on a weed control robot requires an image dataset that shoots plants in real-world conditions.
The picture shows the relative difficulty of species classification.
Most of the current classification methods of weed species also tend to control weeds in planting applications, where it is very simple to classify using machine vision because the land is usually flat and the vegetation is uniform
However, the classification of weeds in the pasture environment is basically ignored.
Ranch environments pose unique challenges to weed management and classification as they are both remote and extensive, with uneven terrain and complex target backgrounds.
In addition, there may be many different kinds of weeds and native plants in the same area, all of which are different distances from the camera, each experiencing different degrees of light and shadow, some weeds are completely hidden.
In order to give any chance of success to the classification methods deployed in this environment, the site-
A specific and highly variable weed species image dataset is required.
Liaise with land conservation groups and owners in northern Australia to select eight target weed species for the collection of large weed species image datasets; (1)chinee apple (), (2)lantana (), (3)parkinsonia (), (4)parthenium (), (5)
Thorn Tree (), (6)rubber vine (), (7)siam weed ()and (8)snake weed (. ).
These species have been selected because they are suitable for spraying leaf-faced herbicides and their notoriety for invasion and destructive effects on rural Australia.
Five of the eight species have been targeted by the Australian government in an attempt to limit their potential transmission and society
Economic impact.
In this study, we present this data set containing 17,509 images of eight different weeds marked by humans.
The images were taken from eight Ranch environments in northern Australia.
In addition, we train a deep learning image classifier to identify species present in each image;
The real-time performance of the classifier is verified.
We expect that the data set and our classification results will stimulate further research into the classification of pasture weeds under realistic conditions, as experienced by autonomous weed control robots.
Custom message
Chat Online 编辑模式下无法使用
Chat Online inputting...