Introduction

Maize cultivar identification is a typical fine-grained visual categorization problem from the view of computer vision. We presented a novel dataset on this problem based on the tassel characteristics of maize and also proposed a deep convolutional neural network (CNN) based recognition framework tailored to the dataset. Note that it is not an easy task to identify the tassel, because many intrinsic and extrinsic variations accompanied in the wild field, including the pose variations, changing illumination, occlusion, blurring, different flowering status, and the cluttered background, would add extra disturbances to an already challenging problem. Indeed, maize cultivar identification does bring new challenges to Computer Vision.

The dataset contains four Chinese maize cultivars of Jundan No.20, Wuyue No.3, Nongda No.108 and Zhengdan No.958 and 5000 tassel images. 10 sequences from 2010 to 2013 are included in constructing the dataset. Concretely, they are Zhengzhou 2010, Zhengzhou 2011, Zhengzhou 2012, Zhengzhou 2013, Taian 2010-1, Taian 2010-2, Taian 2011-1, Taian 2011-2, Taian 2012-1, and Taian 2012-2.

Technical Pipeline

Maize cultivar identification is formulated as a weakly-supervised fine-grained visual categorization problem. To address this, an effective filter-specific feature encoding pipeline is proposed to better encode CNN feature. Several key technologies are included in our recognition framework. They are pretrained deep CNN model, Fisher vector encoding and mutual information based feature selection techniques. The proposed method is tested against compelling and state-of-the-art approaches, and it exhibits superior performance by at least 5% in accuracy.

Filter-Specific Encoding Mechanism

The activations of filtermaps are used as identity maps to guide which column features should be aggregated.

Mutual Information Based Filter Selection

We observe that not all filters are essential for classification. Some irrelevant filters are discarded accorading to the mutual information between representations and labels.

Dataset

As promised, our maize cultivar identification dataset termed HUST-FG-MCI is released here. Notice that there are not that many agriculture-orientated literatures and methods coming from Computer Vision community applied specifically to address those in-field visual challenges. The main reason perhaps is that there lacks suitable and publicly available datasets to allow solid scientific works. However, it does take time to collect images in agriculture, because we cannot break the natural rule of crop growth. Step by step, little by little, we tend to break such a wall so that other talented and dedicated scientists could involve in this virgin field.

To download the the HUST-FG-MCI dataset, you agree that:

• the dataset will be used for academic purposes only
• you will cite this paper if you use the dataset in your publication
• you will not distribute the dataset to others
• you must ask permission of the corresponding author if the dataset will be used for commercial purposes

HUST-FG-MCI

• Google Drive

Code

We provide here a complete Matlab implementation of our approach decribed in our paper. One could use this code to reproduce exactly the experimental results. Also, the implementations of several comparing methods are also included.

Notice that one needs to download pretrained models at MatConvNet homepage (http://www.vlfeat.org/matconvnet/) or download fine-tuned versions through the links below.

Source code can be downloaded from:

MCI-Matlab-release-v1.0 Google Drive Baidu Yun

Fine-tuned models can be download from:

imagenet-vgg-m-1024-fine-tuned Google Drive Baidu Yun

imagenet-vgg-verydeep-16-fine-tuned Google Drive Baidu Yun

Publication

Hao Lu, Zhiguo Cao, Yang Xiao, Zhiwen Fang, Yanjun Zhu, Fine-grained maize cultivar identification using filter-specific convolutional activations, IEEE International Conference on Image Processing (ICIP), pages 3718-3722, 2016.

Hao Lu, Zhiguo Cao, Yang Xiao, Zhiwen Fang, Yanjun Zhu, Toward Good Practices for Fine-Grained Maize Cultivar Identification With Filter-Specific Convolutional Activations, IEEE Transactions on Automation Science and Engineering, volume 15, number 2, pages 430-442, 2018.

Citation

Conference Version

@InProceedings{lu2016fine,

  author    = {Lu, Hao and Cao, Zhiguo and Xiao, Yang and Fang, Zhiwen and Zhu, Yanjun},

  title     = {Fine-grained maize cultivar identification using filter-specific convolutional activations},

  booktitle = {Proc. IEEE International Conference on Image Processing (ICIP)},

  year      = {2016},

  pages     = {3718--3722},

  doi       = {10.1109/ICIP.2016.7533054},

}

Journal Version

@Article{lu2018toward,

  author  = {Lu, Hao and Cao, Zhiguo and Xiao, Yang and Fang, Zhiwen and Zhu, Yanjun},

  title   = {Toward Good Practices for Fine-Grained Maize Cultivar Identification With Filter-Specific Convolutional Activations},

  journal = {IEEE Transactions on Automation Science and Engineering},

  volume  = {15},

  number  = {2},

  pages   = {430--442},

  year    = {2018},

  doi     = {10.1109/TASE.2016.2616485},

}