It is a challenging task to recognize smoke from images due large variance of color, texture, and shapes. There are detection methods that have been proposed, but most them based on hand-crafted features. To improve the performance detection, we propose novel deep normalization convolutional neural network (DNCNN) with 14 layers implement automatic feature extraction classification. In DNCNN, traditional replaced accelerate training process boost detection. reduce overfitting caused by...

Anomaly detection in surveillance videos is a challenging task due to the diversity of anomalous video content and duration. In this paper, we consider anomaly as regression problem with respect scores clips under weak supervision. Hence, propose an framework, called Regression Net (ARNet), which only requires video-level labels training stage. Further, learn discriminative features for detection, design dynamic multiple-instance learning loss center proposed AR-Net. The former used enlarge...

Abstract Anomaly detection has attracted considerable search attention. However, existing anomaly databases encounter two major problems. Firstly, they are limited in scale. Secondly, training sets contain only video‐level labels indicating the existence of an abnormal event during full video while lacking annotations precise time durations. To tackle these problems, we contribute a new L arge‐scale A nomaly D etection ( LAD ) database as benchmark for sequences, which is featured aspects....

Autism spectrum disorder (ASD) is a common mental illnesses for children. Existing studies show that visual attention of children with ASD different from normal Thus, it meaningful to design an effective model diagnose ASD. In this paper, we propose saliency detection by small-scale image samples. the proposed model, deep neural network high-level semantic features in data-driven way. The U-net used construct feature learning, where loss function positive and negative equilibrium mean...

Attention supervision encourages grounded video description models (GVDMs) to focus on the related visual content when generating words. Thus, it improves performance of GVDMs. However, existing GVDMs often fail small but informative regions because these are considered as negative by using intersection-over-union (IoU) based attention groundtruth sampling method. Moreover, prevailing loss functions enforce equally all sampled generate words, which may make difficult for model attend and...

Self-attention based encoder-decoder models achieve dominant performance in image captioning. However, most existing captioning (ICMs) only focus on modeling the relation between spatial tokens, while channel-wise attention is neglected for getting visual representation. Considering that different channels of representation usually denote objects, it may lead to poor terms object and attribute words sentences generated by ICMs. In this paper, we propose a novel dual-stream self-attention...

Inspired by the recent success of fully convolutional networks (FCN) in semantic segmentation, we propose a deep smoke segmentation network to infer high quality masks from blurry images. To overcome large variations texture, color and shape appearance, divide proposed into coarse path fine path. The first is an encoder-decoder FCN with skip structures, which extracts global context information accordingly generates mask. retain spatial details smoke, second also designed as but it shallower...

Anomaly detection in surveillance videos is a challenging task due to the diversity of anomalous video content and duration. In this paper, we consider anomaly as regression problem with respect scores clips under weak supervision. Hence, propose an framework, called Regression Net (AR-Net), which only requires video-level labels training stage. Further, learn discriminative features for detection, design dynamic multiple-instance learning loss center proposed AR-Net. The former used enlarge...

Anomaly detection has attracted considerable search attention. However, existing anomaly databases encounter two major problems. Firstly, they are limited in scale. Secondly, training sets contain only video-level labels indicating the existence of an abnormal event during full video while lacking annotations precise time durations. To tackle these problems, we contribute a new Large-scale Detection (LAD) database as benchmark for sequences, which is featured aspects. 1) It contains 2000...