Search

In this paper, we propose a novel deep learning-based feature learning architecture for object classification. Conventionally, deep learning methods are trained with supervised learning for object classification. But, this would require large amount of training data. Currently there are increasing trends to employ unsupervised learning for deep learning. By doing so, dependency on the availability of large training data could be reduced. One implementation of unsupervised deep learning is for feature learning where the network is designed to “learn” features automatically from data to obtain good representation that then could be used for classification. Autoencoder and generative adversarial networks (GAN) are examples of unsupervised deep learning methods. For GAN however, the trajectories of feature learning may go to unpredicted directions due to random initialization, making it unsuitable for feature learning. To overcome this, a hybrid of encoder and deep convolutional generative adversarial network (DCGAN) architectures, a variant of GAN, are proposed. Encoder is put on top of the Generator networks of GAN to avoid random initialisation. We called our method as EGAN. The output of EGAN is used as features for two deep convolutional neural networks (DCNNs): AlexNet and DenseNet. We evaluate the proposed methods on three types of dataset and the results indicate that better performances are achieved by our proposed method compared to using autoencoder and GAN.

It is well-known that a large amount of data is required to train deep learning systems. However, data collection is very costly if it is not impossible to do. To overcome the limited data problem, one can use models that have been trained with a large dataset and apply them in the target domain with a limited dataset. In this paper, we use pre-trained models on imageNet data and re-train them on our data to detect tea leaf diseases. Those pre-trained models use deep convolutional neural network (DCNN) architectures: VGGNet, ResNet, and Xception. To mitigate the difference tasks of ImageNet and ours, we apply fine-tuning on the pre-trained models by replacing some parts of the pre-trained models with new structures. We evaluate the performance using various re-training and fine-tuning schema. The vanilla pre-trained model is used as the baseline while other techniques such as re-training the models on the appended structures, partially re-training the pre-trained models, and fully re-training the whole networks where the pre-trained models are used in the initialization as the evaluator. Our experiments show that applying transfer learning only on our data may not be effective due to the difference in our task to ImageNet. Applying fine-tuning on pre-trained DCNN models is found to be effective. It is consistently better than that of using transfer learning only or partial fine-tuning. It is also better than training the model from scratch, i.e., without using pre-trained models.

At the moment, there are increasing trends of using deep learning for plant diseases detection. However, their implementations may be difficult in developing countries due to several reasons. First, existing deep learning models are usually trained with images with adequate resolutions. In developing countries however, with limited internet connection, models that would perform well even when data with low resolution are used are needed. Secondly, the generated models are large. Hence, most deep learning based applications are available on-line. Unfortunately, the trend for new deep learning architectures are either have larger models or require a heavy memory usage. So, models with smaller size would be preferred. In this paper, we evaluate various existing deep learning models for plant diseases detection when low resolution data are used. They are: VGGNet, AlexNet, Resnet, Xception, and MobileNet. Our focus is deep convolutional neural network (DCNN) which is commonly applied for image data. We also propose a new DCNN architecture with two branches of concatenated residual networks. It is well known that the deeper the networks the better performance of DCNN. However, DCNN with very deep networks and large number of training parameters is prone to vanishing gradient problems. One solutions for that is to apply residual networks as branches to DCNN. While it is found that increasing the branch of the networks benefit the performance, larger memory are required to train the networks. So, we apply two concatenated residual networks only. We called it Compact Networks (ComNet). We compare our method other with six popular CNN architectures. We evaluate the performance on the PlantVillage dataset and our own dataset. We collected images of tea leaves which consist of 6 classes: 5 classes of diseases that are commonly found in Indonesia and a healthy class. Our experiments show that our method is generally better than referenced DCNN networks.

Traditionally, automatic speech recognition (ASR) uses a Hidden Markov Model with Gaussian Mixture Model (HMM-GMM) as acoustic model and hand-designed features such as Mel-frequency Cepstral Coefficient (MFCC) as acoustic features. It is usually assumed that the features are uncorrelated, making it possible to use diagonal covariances for the GMM. The assumption generally holds due to the use of Discrete Cosine Transformation (DCT) that de-correlates the speech spectra. However, DCT could cause some information loss, such as correlations between the feature components. Current ASR systems, which is based on Deep Neural Network (DNN) show to be better especially in reverberant conditions when more primitive features, such as filter-bank (FBANK), are used. This might be because DNN is better in modeling non-linear relations between the components of the features. But the use of short-time processing in FBANK may cause the lost of long-term correlations in a speech pattern. To tackle this, we propose a new feature, q-FBANK, which is a generalization of FBANK. The results on artificially reverberant speech show that the proposed features achieve better performance than MFCC and FBANK on DNN-HMM systems where an average error reduction up to 39.73% and 13.5% were achieved respectively.

Early detection is critical for maintaining quantity and quality of farming commodity. Currently, detection of plant diseases still requires human expertise and/or need microscopic identification such as spectroscopic technique and molecular biological. So, it would be very costly and time consuming, and hence unattainable for small-holder farmers. The rapid development of intelligent agriculture using machine learning has led the widespread use of computer or smart-phones to solve this problem. So, early detection of plant disease can be performed with minimal support from human experts and microscopic identification is no longer needed. However, conventional machine-learning techniques are limited in their ability to process raw data directly. So it require some efforts and domain expertise to design feature extractor to support it. Moreover, impulse noise such as salt-pepper noise may present on the images and it arises another challenge to provide a robust system. In this paper, we present denoising convolutional variational autoencoders as automatic unsupervised feature extractor and automatic denoiser to learn and to extract good features directly from the raw data. Here, we use the output of denoising convolutional variational auto encoders as inputs to fully connected networks classifiers for automatic detection of plant diseases. Our experiments show the average accuracies of our method is better than denoising variational autoencoders which is built using fully deep connected networks architectures. We also found that our proposed method is more robust against noisy test data.

Deep learning technology has a better result when trained using an abundant amount of data. However, collecting such data is expensive and time consuming. On the other hand, limited data often be the inevitable condition. To increase the number of data, data augmentation is usually implemented. By using it, the original data are transformed, by rotating, shifting, or both, to generate new data artificially. In this paper, generative adversarial networks (GAN) and deep convolutional GAN (DCGAN) are used for data augmentation. Both approaches are applied for diseases detection. The performance of the tea diseases detection on the augmented data is evaluated using various deep convolutional neural network (DCNN) including AlexNet, DenseNet, ResNet, and Xception. The experimental results indicate that the highest GAN accuracy is obtained by DenseNet architecture, which is 88.84%, baselines accuracy on the same architecture is 86.30%. The results of DCGAN accuracy on the use of the same architecture show a similar trend, which is 88.86%.

One factor affecting the quality of tea is the selection of plant material that would be planted on the field. Clonal selection is a common way to produce tea with better quality. However, as a natural cross pollination species, tea often consists of various clones or progenies of cross-pollinated process. This commonly occurs on plantations owned by smallholder farmers. To produce a consistent quality tea, the clones or progenies need to be identified. Usually, human experts distinguish the plants from leaves by visual inspection on the physical attributes of the leaves, such as the textures, the bone structures, and the colors. It is very difficult for non-experts or common farmers to do such identifications. In this, we propose a deep learning-based identification of tea clones. We apply deep convolutional neural network (CNN) to identify 3 types of tea clones of Gambung series, a series of tea clones developed at Research Institute of Tea and Cinchona. Our study indicates that the performance of the CNN systems are affected by the depth of the convolutional layers. VGGNet, a popular CNN architectures with 16 layers, achieves better accuracy compared to AlexNet, a CNN with 6 layers.