Ye Luo

Deep Salient Object Detection via Hierarchical Network Learning

Salient object detection is a fundamental problem in both pattern recognition and image processing tasks. Previous salient object detection algorithms usually involve various features based on priors/assumptions about the properties of the objects. Inspired by the effectiveness of recently developed feature learning, we propose a novel deep salient object detection (DSOD) model using the deep residual network (ResNet 152-layers) for saliency computation. In particular, we model the image saliency from both local and global perspectives. In the local feature estimation stage, we detect local saliency by using a deep residual network (ResNet-L) which learns local region features to determine the saliency value of each pixel. In the global feature extraction stage, another deep residual network (ResNet-G) is trained to predict the saliency score of each image based on the global features. The final saliency map is generated by a conditional random field (CRF) to combining the local and global-level saliency map. Our DSOD model is capable of uniformly highlighting the objects-of-interest from complex background while well preserving object details. Quantitative and qualitative experiments on three benchmark datasets demonstrate that our DSOD method outperforms state-of-the-art methods in the salient object detection.

MC-DCNN: Dilated Convolutional Neural Network for Computing Stereo Matching Cost

Designing a model for computing better matching cost is a fundamental problem in stereo method. In this paper, we propose a novel convolutional neural network (CNN) architecture, which is called MC-DCNN, for computing matching cost of two image patches. By adding dilated convolution, our model gains a larger receptive field without adding parameters and losing resolution. We also concatenate the features of last three convolutional layers as a better descriptor that contains information of different image levels. The experimental results on Middlebury datasets validate that the proposed method outperforms the baseline CNN network on stereo matching problem, and especially performs well on weakly-textured areas, which is a shortcoming of traditional methods.

Image salient object detection with refined deep features via convolution neural network

Abstract. Recent advances in saliency detection have used deep learning to obtain high-level features to detect salient regions. These advances have demonstrated superior results over previous works that use handcrafted low-level features for saliency detection. We propose a convolutional neural network (CNN) model to learn high-level features for saliency detection. Compared to other methods, our method presents two merits. First, when performing features extraction, apart from the convolution and pooling step in our method, we add restricted Boltzmann machine into the CNN framework to obtain more accurate features in intermediate step. Second, in order to avoid manual annotation data, we add deep belief network classifier at the end of this model to classify salient and nonsalient regions. Quantitative and qualitative experiments on three benchmark datasets demonstrate that our method performs favorably against the state-of-the-art methods.

3D Spatial Pyramid Dilated Network for Pulmonary Nodule Classification

Lung cancer mortality is currently the highest among all kinds of fatal cancers. With the help of computer-aided detection systems, a timely detection of malignant pulmonary nodule at early stage could improve the patient survival rate efficiently. However, the sizes of the pulmonary nodules are usually various, and it is more difficult to detect small diameter nodules. The traditional convolution neural network uses pooling layers to reduce the resolution progressively, but it hampers the network’s ability to capture the tiny but vital features of the pulmonary nodules. To tackle this problem, we propose a novel 3D spatial pyramid dilated convolution network to classify the malignancy of the pulmonary nodules. Instead of using the pooling layers, we use 3D dilated convolution to learn the detailed characteristic information of the pulmonary nodules. Furthermore, we show that the fusion of multiple receptive fields from different dilated convolutions could further improve the classification performance of the model. Extensive experimental results demonstrate that our model achieves a better result with an accuracy of 88.6%, which outperforms other state-of-theart methods.

An Adversarial and Densely Dilated Network for Connectomes Segmentation

Automatic reconstructing of neural circuits in the brain is one of the most crucial studies in neuroscience. Connectomes segmentation plays an important role in reconstruction from electron microscopy (EM) images; however, it is rather challenging due to highly anisotropic shapes with inferior quality and various thickness. In our paper, we propose a novel connectomes segmentation framework called adversarial and densely dilated network (ADDN) to address these issues. ADDN is based on the conditional Generative Adversarial Network (cGAN) structure which is the latest advance in machine learning with power to generate images similar to the ground truth especially when the training data is limited. Specifically, we design densely dilated network (DDN) as the segmentor to allow a deeper architecture and larger receptive fields for more accurate segmentation. Discriminator is trained to distinguish generated segmentation from manual segmentation. During training, such adversarial loss function is optimized together with dice loss. Extensive experimental results demonstrate that our ADDN is effective for such connectomes segmentation task, helping to retrieve more accurate segmentation and attenuate the blurry effects of generated boundary map. Our method obtains state-of-the-art performance while requiring less computation on ISBI 2012 EM dataset and mouse piriform cortex dataset.

A Hybrid Model: DGnet-SVM for the Classification of Pulmonary Nodules

We investigate the problem of benign and malignant pulmonary nodules classification for thoracic Computed Tomography (CT) images. Although various methods have been proposed to solve this problem, they have bottlenecks of poor input image quality and subjective or shallow feature extraction. In this paper, we propose a Denoise GoogLeNet model with the classifier of Support Vector Machine (DGnet-SVM) to improve the final classification accuracy. We apply Denoise Network to improve the CT image quality by reducing the noise, and GoogLeNet is utilized to extract high-level features for better generalization of data. Furthermore, SVM is applied to classify the nodules owing to its great classification performance. The experimental results show that our hybrid model outperforms other state-of-the-art methods with the accuracy of 0.89 based on five-fold cross validation and the AUC is 0.95. The advantages of the proposed model and our future work are also discussed.

Regularizing CNN via Feature Augmentation

Very deep convolutional neural network has a strong representation power and becomes the dominant model to tackle very complex image classification problems. Due to the huge number of parameters, overfitting is always a primary problem in training a network without enough data. Data augmentation at input layer is a commonly used regularization method to make the trained model generalize better. In this paper, we propose that feature augmentation at intermediate layers can be also used to regularize the network. We implement a modified residual network by adding augmentation layers and train the model on CIFAR10. Experimental results demonstrate our method can successfully regularize the model. It significantly decreases the cross-entropy loss on test set although the training loss is higher than the original network. The final recognition accuracy on test set is also improved. In comparison with Dropout, our method can cooperate better with batch normalization to produce performance gain.

Scanpath Prediction Based on High-Level Features and Memory Bias

Human scanpath prediction aims to use computational models to mimic human gaze shifts under free view conditions. Previous works utilizing low-level features, hand-crafted high-level features, saccadic amplitude, memory bias cannot fully explain the mechanism of visual attention. In this paper, we propose a comprehensive method to predict scanpath from four aspects: low-level features, saccadic amplitude, semantic features learned via deep convolutional neural network, memory bias including short-term and long-term memory. By calculating the probabilities for all candidate regions in an image, the position of next fixation point can be selected via picking the one with the largest probability product. Moreover, fixation duration as a key factor is first used to model memory effect on scanpath prediction. Experiments on two public datasets demonstrate the effectiveness of the proposed method, and comparisons with state-of-the-art methods further validate the superiority of our method.