Zhi-Kai Huang

Hybrid Deep Learning for Plant Leaves Classification

Recently, deep learning is very popular, it has been applied into many applications, In this paper, a new neural network, hybrid deep learning is introduced, which included AutoEncoder(AE) and convolutional neural network (CNN). This neural network is applied for extracting the features of the plant leaves. In this paper, we proved that hybrid deep learning can extract better features for classification task. We apply the hybrid deep learning to extract features of leaf pictures, and then we classify leaves using those features with SVM, the result suggests that this method is not only better than pure SVM, but also better than pure AE and pure CNN.

Convolutional Neural Network Application on Leaf Classification

Plants are everywhere in our lives, we can classify them by observing their features. But for ordinary people, the species we don’t know are much more than we know. So, for amateurs who are interested in botany, a system which can classify different species of leaves must be very useful, a system like that will also help students recognize the leaves they don’t know. This paper describes a system for leaf classification, which is developed with convolutional neural network technique. Previous researches in leaf identification usually use grayscale images. The main reason is that these samples mostly are green leaves. This system is trained by 1500 leaves to classify 50 kinds of plants. Compared to other research, our net use RGB images for input. And in convolutional neural network, we use PReLU instead of traditional ReLU. The experimental result shows that our method for classification gives accuracy of 94.8 %.

Leaf Classification Utilizing a Convolutional Neural Network with a Structure of Single Connected Layer

Plant plays an important role in human life, so it is necessary to build an automatic system for recognizing plant. Leaf classification has become a research focus for twenty years. In this paper, we propose a single connected layer (SCL) structure adding into the convolutional neural network (CNN). We use this CNN model for plant leaf identification and report the promising results on ICL leaf database. Moreover, we propose some improvement on it to let it perform better. The result shows that our advanced SCL can effectively improve the accuracy of CNN.

Chinese Historic Image Threshold Using Adaptive K-means Cluster and Bradley’s

Resorting to extraction text techniques for Chinese heritage documents becomes an increasing need. Historic documents such as Chinese calligraphy usually were handwritten or scanned in low contrast so that an automatic optical character recognition procedure for document images analysis is difficult to apply. In this paper, we present a historic document image threshold based on a combination of Bradley’s algorithm and K-means. An adaptive K-means cluster as a pre-processing methods for document image has been used for automatically grouping the pixels of a document image into different homogeneous regions. In Bradley’s methods, every image’s pixel is set to black if its brightness is T percent lower than the average brightness of surrounding pixels in the window of the specified size, otherwise it is set to white. Finally, text bounding boxes are generated by concatenating neighboring word clusters with mathematical morphology method. Experimental results show that this algorithm is robust in dealing with non-uniform illuminated, low contrast historic document images in terms of both accuracy and efficiency.

An Adaptive Segmentation Algorithm for Degraded Chinese Rubbing Image Binarization Based on Background Estimation

Image Segmentation plays an important role in image processing and analysis. In order to preserve strokes of a Chinese character while enhancing character details for degraded historical document image, we propose an adaptive segmentation algorithm for degraded historical document image binarization based on background estimation for non-uniform illumination images. The novelty of the proposed method is that find an optimal background estimation based on Blind/Referenceless Image Spatial QUality Evaluator. The proposed method has four steps: (i) preprocess using median filtering; (ii) extraction of the red color components; (iii) a morphological operation in order to find an optimal background estimation; and (iv) segmented binary image using Otsu’s Thresholding. Experimental results demonstrate that it is capable of extracting more accurate segmentation of characters for degraded Chinese rubbing document image.

Visual data completion via local sensitive low rank tensor learning

The problems of estimating missing values in visual data appear ubiquitously in computer vision applications including image inpainting, video inpainting, hyperspectral data recovery, and magnetic resonance imaging (MRI) data recovery. Recently, it is shown that tensor completion, which generalizes matrix completion to multiway data of higher order, could accurately estimate the overall data structure and achieves state-of-the-art performance for video completion. However, current tensor-based approaches implicitly assume that the partially observed video is globally low rank, which is too stringent for practical applications where the input video could include multiple heterogeneous episodes and the global correlation between frames is not high. To tackle this problem, we propose a novel local sensitive formulation of tensor learning where we assume instead that the video is inter-correlated in a local manner, leading to a representation of the observed tensor as a weighted sum of low-rank tensors. Computationally, we also design efficient scheme for solving the resulting learning problem based on the alternating direction method of multipliers (ADMM). Our experiments show improvements in prediction accuracy over classical approaches for visual data completion tasks.

Locally Biased Discriminative Clustering Method for Interactive Image Segmentation

Interactive image segmentation is a form of semi-supervised segmentation method by using the user interactive information. It performed well than fully unsupervised segmentation methods. In this paper, we propose a novel interactive image segmentation method, in which a seed vector is used to represent the user scribbles. Then a soft similarity constraint is added to the discriminative clustering model. The soft constraint allows the user to tune the degree which the constraint is satisfied. With respect to the discriminative clustering model, the clustering result is not affected by the assumption to the distribution of the data, and it’s easy to add constraint to the clustering variable. The final optimization problem is convex, so it can reach global optimal solution. The proposed method is evaluated on benchmark dataset BSD dataset, and it performs well than state of art methods both in quantitative and qualitative results.

Implementation of Plant Leaf Recognition System on ARM Tablet Based on Local Ternary Pattern

The Local Binary Pattern (LBP) and its variants is powerful in capturing image features and computational simplicity, However LBP’s sensitivity to noise, particularly in near-uniform image regions has stimulated many transformations of LBP to improve the ability of feature description. The Local Ternary Pattern (LTP) extends the conventional LBP to ternary codes and makes a significant improvement. LTP is more resistant to noise, but no longer strictly invariant to gray-level transformations. In this paper, by adopting the Average Local Gray Level (ALG) to take place of the traditional gray value of the center pixel and taking an auto-adaptive strategy on the selection of the threshold, we propose the Enhanced Local Ternary Pattern (ELTP) to improve the performance of LTP and implement an android application to recognize plant-leaf image and identify the species of the plant.

Binarization Chinese Rubbing Images Using Gaussian Mixture Model

Rubbings are important components of ancient Chinese books, and are the main source for people to learn, study, and research history. Image segmentation plays a crucial role in extracting useful information and characteristics of Chinese character from the rubbing images. In this paper, binarization using a Gaussian Mixture Model (GMM) with 2 components for representation of background and foreground distribution in a Chinese rubbing image has been proposed. To model the likelihood of each pixel belonging to foreground or background, a foreground and background color model are learned from three color bands samples that using RGB color space. The standard Expectation-Maximisation (EM) algorithm had been used to estimate the GMM parameters. Experimental results on real rubbing images validate the effectiveness of the model when working with Chinese rubbing images.

Plant Leaf Recognition Based on Contourlet Transform and Support Vector Machine

Plants are essential to the balance of nature and in people’s lives as the fundamental provider for food, oxygen and energy. The study of plants is also essential for environmental protection and helping farmers increase the production of food. As a fundamental task in botanical study, plant leaf recognition has been a hot research topic in these years. In this paper, we propose a new method based on contourlet transform and Support Vector Machine (SVM) for leaf recognition. Contourlet Transform is a promising multi-resolution analysis technique, which provides image with a flexible anisotropy and directional expansion. By basing its constructive principle on a non-subsampled pyramid structure and related directional filter banks, contourlet transform decomposes input images into multi-scale factors which also enjoys additional advantages such as shift invariance and computational efficiency. Compared with one-dimensional transforms, such as the Fourier and wavelet transforms, Contourlet Transform can capture the intrinsic geometrical structure. In order to ameliorate the influence of unwanted artefacts such as illumination and translation variations, in this paper, the contourlet transform was firstly applied to extract feature with high discriminative power. Then the extracted features are classified by SVM. The experimental results show that the proposed method has high sensitivity of directionality and can better capture the rich features of natural images such as edges, curves and contours.