Automatic recognition of facial expression images is a challenge for computer due to variation of expression, background, position and label noise. The paper propose a new method for static facial expression recognition. Main process is to perform experiments by FER-2013 dataset, the primary mission is using our CNN model to classify a set of static images into 7 basic emotions and then achieve effective classification automatically. The two preprocessing of the faces picture have enhanced the effect of the picture for recognition. First, FER datasets are preprocessed with standard histogram eqialization. Then we employ ImageDataGenerator to deviate and rotate the facial image to enhance model robustness. Finally, the result of softmax activation function (also known as multinomial logistic regression) is stacked by SVM. The result of softmax activation function + SVM is better than softmax activation function. The accuracy of facial expression recognition achieve 68.79% on the test set.
Automatic image cropping is of great importance for improving visual quality of images. Existing methods typically repurpose classifiers to perform cropping. Instead, we frame image cropping as a regression problem to bounding boxes and associated visual quality scores. An end-to-end framework with fully convolutional network is proposed to predict bounding boxes and associated scores from full images in one evaluation, which accepts input images of arbitrary sizes. In this way, we circumvent the requirement of fixed-sized inputs from prevalent convolutional networks, and thus avoid impairing the visual quality of images. Additionally, the whole cropping pipeline is a single network, so it can be optimized end-to-end directly on cropping performance, which is different from existing methods requiring the processes of proposal generation and subsequent region feature extraction. Experiments on a public dataset confirm that our framework has competitive accuracy and speed to traditional approaches.
Automatic makeup aims to compose and synthesize makeup on human face through computer algorithm, which belongs to the field of face image analysis. It plays an important role in interactive entertainment applications, image and video editing and face recognition assistance. However, as a face editing task, automatic makeup has several challenges as following: 1) It is usually difficult to meet the editing requirements while ensuring that makeup results are naturally looking without obvious artifacts. 2) Designed network should preserve the content (human identity) of input images but only edits the makeup-related parts of input images such as lip gloss and eye shadow. 3) Number of face makeup dataset is limited. To address these difficulties, we propose an automatic makeup generative adversarial network, a novel approach that can perform makeup attribute editing for multiple makeup styles. The network is an end-to-end model with several specially designed skip layers in generator network and allows simultaneous training of multiple datasets in addition to makeup datasets. Finally, the performance of proposed method was compared with state-of-art method qualitatively and quantitatively.
Traditional text classification methods make a basic assumption: the training and test data are homologous, while this naive assumption may not hold in the real world. Hence, this paper studies the problem of domain-adapted news text classification, hereby a model is trained on labeled data from one source domain and is able to be deployed on the other. To realize the cross-domain text classification, we propose a domain-adapted text classification method based on topic model LDA and TextCNN model, named TextLDACNN. Specifically, our work calculates the topic similarity between source and target domain, which is severed as an effective constraint to regularize the training process and hence improve the generalization of the source model to the target domain. Text classifier trained with unsupervised topic feature representation clearly outperforms the baseline TextCNN model. The result shows that our method achieves an approximately 4.0% improvement compared to the state-of-the-art method.
With the development of cloud computing and big data, more and more applications are deployed in the cloud. Data center, as a management unit of computing and storage resources, consists of quantities of computing clusters, and the scale of clusters will be increased as demand grows. Even the same application will have a totally different impact on the network load when it is under clusters with different sizes, for the topology of the clusters is changed. Therefore, it is meaningful to do research on how to make use of the load of network from an application under a small cluster to predict the network load under a larger cluster, which will guide the data center on its infrastructure. In this paper, first of all, we address a new network load forecasting problem and collect network load generated by the operations of a designated computing intensive task (Terasort and GroupBy). Then we analyze the data and put forward the difficulties of the new problem. Finally, with a given set of data, we test the traditional machine learning models and the deep learning models respectively and provide a simple but effective model.
