Selective feature connection mechanism: Concatenating multi-layer CNN features with a feature selector

Exploiting the complementary strengths of multi-layer CNN features for image retrieval

Pre-trained DCNN trained on a large-scale image database can be used as a universal feature representation for image classification, which has achieved significant progress in some image recognition tasks. Compared with other image recognition tasks, directly utilizing a single convolutional feature as feature representation for vein recognition task cannot achieve the impressive result due to the sparse distribution of vein information. Therefore, to obtain more representative and discriminative convolutional feature for vein recognition, a novel multi-layer convolutional features' concatenation with semantic feature selector is proposed in this paper. In the pre-trained DCNN, different convolutional layers can encode different-level feature information. High-level convolutional features with vein information cover more semantic information and low-level convolutional features with vein information cover more detail information. However, low-level convolutional features also contain some background information. Therefore, in order to remove the background information of low-level convolutional features, a novel semantic feature selector is presented. First, the proposed local max-pooling of preserving spatial position (LMP-PSP) information is applied on activation map obtained by adding up all feature maps of the high-level convolutional layer to generate the semantic weighting map, which reflects key vein information of high-level convolutional features. Then, semantic weighting map is regarded as a feature selector to discard the background information of the low-level convolutional features and preserve the detail information of low level convolutional features. Finally, low-level convolutional features with vein information are selectively linked to high-level convolutional features with vein information based on the proposed semantic feature selector. A series of rigorous experiments on two lab-made vein databases named CUMT-Hand-Dorsa Vein database and CUMT-Palm Vein database is conducted to verify the effectiveness and feasibility of the proposed model. Besides, additional experiments with PUT Palm Vein database and the subset of PolyU database illustrate its generalization ability and robustness.

Disaster damage detection through synergistic use of deep learning and 3D point cloud features derived from very high resolution oblique aerial images, and multiple-kernel-learning

Recent works about convolutional neural networks (CNN) show breakthrough performance on various tasks. However, most of them only use the features extracted from the topmost layer of CNN instead of leveraging the features extracted from different layers. As the first group which explicitly addresses utilizing the features from different layers of CNN, we propose cross-layer CNN features which consist of the features extracted from multiple layers of CNN. Our experimental results show that our proposed cross-layer CNN features outperform not only the state-of-the-art results but also the features commonly used in the traditional CNN framework on three tasks — artistic style, artist, and architectural style classification. As shown by the experimental results, our proposed cross-layer CNN features achieve the best known performance on the three tasks in different domains, which makes our proposed cross-layer CNN features promising solutions for generic tasks.

This paper proposes an efficient and robust Loop closure detection (LCD) method based on Convolutional neural network (CNN) feature. The primary method is called SeqCNNSLAM, in which both the outputs of the intermediate layer of a pre-trained CNN and the outputs of traditional sequence-based matching procedure are incorporated, making it possible to handle the viewpoint and condition variance properly. An acceleration algorithm for SeqCNNSLAM is developed to reduce the search range for the current image, resulting in a new LCD method called A-SeqCNNSLAM. To improve the applicability of A-SeqCNNSLAM to new environments, O-SeqCNNSLAM is proposed for online parameters adjustment in A-SeqCNNSLAM. In addition to the above work, we further put forward a promising idea to enhance SeqSLAM by integrating the both CNN features and VLAD's advantages called patch based SeqCNNSLAM (P-SeqCNNSLAM), and provide some preliminary experimental results to reveal its performance.

Correlation tracker has made a huge success in visual object tracking. However, it is mainly because that the tracker cannot catch the occurrence of appearance changes, tracking based on correlation filters often drifts due to the unexpected appearance changes caused by occlusion, deformation and background clutter. In this paper, we propose a new method to detect the case when the tracker encountered the unexpected appearance changes. This method uses the following points: 1) Filter response curve would decreases dramatically when target suffers heavy appearance changes. 2) Features extracted from deeper layers of convolutional neural networks (CNNs) have more semantics information and features extracted from shadower layers have more spatial information. Extensive experimental results on several public benchmark datasets show that the proposed method can deal with the appearance changes effectively.

How to properly involve text characteristics like multi-scale, arbitrary direction, length aspect ratio, into detection network design has become a hot topic in computer vision. Feature Pyramid Network (FPN) is a typical method to achieve robust text detection, where its low-level and high-level feature map retains spatial structure and global semantic information, respectively. However, its strict hierarchical structure fails to fuse low-level and high-level information to improve the distinguish ability of feature map. To address this problem, we propose a novel feature fusion pyramid network for end-to-end scene text detection by fusing multi-modal information. By dividing pyramid structure into high-level and low-level layers, channel and spatial attention modules are adopted to enhance high-level and low-level feature representation by encoding channel and spatial-wise context information, respectively. In order to reduce information loss by layer transmission, a special residual network is designed to achieve short-cut between high-level and low-level features, so as to realize multi-modal feature fusion. Experiments show the precision and recall of the proposed method on ICDAR2015, ICDAR2017-MLT, and MSRA-TD500 datasets reach 88.7%/82.1%, 77.0%/60.3%, and 85.3%/74.8%, respectively.

Deep convolutional neural networks (CNNs) have been widely and successfully applied in many computer vision tasks, such as classification, detection, semantic segmentation, and so on. As for image retrieval, while off-the-shelf CNN features from models trained for classification task are demonstrated promising, it remains a challenge to learn specific features oriented for instance retrieval. Witnessing the great success of low-level SIFT feature in image retrieval and its complementary nature to the semantic-aware CNN feature, in this paper, we propose to embed the SIFT feature into the CNN feature with a Siamese structure in a learning-based paradigm. The learning objective consists of two kinds of loss, i.e., similarity loss and fidelity loss. The first loss embeds the image-level nearest neighborhood structure with the SIFT feature into CNN feature learning, while the second loss imposes that the CNN feature with the updated CNN model preserves the fidelity of that from the original CNN model solely trained for classification. After the learning, the generated CNN feature inherits the property of the SIFT feature, which is well oriented for image retrieval. We evaluate our approach on the public data sets, and comprehensive experiments demonstrate the effectiveness of the proposed method.

Recently, the Siamese network based visual tracking methods have shown great potentials in balancing the tracking accuracy and computational efficiency. These methods use two-branch convolutional neural networks (CNNs) to generate a response map between the target exemplar and each of candidate patches in the search region. However, since these methods have not fully exploit the target-specific information contained in the CNN features during the computation of the response map, they are less effective to cope with target appearance variations and background clutters. In this paper, we propose a Target-Specific Response Attention (TSRA) module to enhance the discriminability of these methods. In TSRA, a channel-wise cross-correlation operation is used to produce a multi-channel response map, where different channels correspond to different semantic information. Then, TSRA uses an attention network to dynamically re-weight the multi-channel response map at every frame. Moreover, we introduce a shortcut connection strategy to generate a residual multi-channel response map for more discriminative tracking. Finally, we integrate the proposed TSRA into the classical Siamese based tracker (i.e., SiamFC) to propose a new tracker (called TSRA-Siam). Experimental results on three popular benchmark datasets show that the proposed TSRA-Siam outperforms the baseline tracker (i.e., SiamFC) by a large margin and obtains competitive performance compared with several state-of-the-art trackers.

A pooling-based feature pyramid network for salient object detection

In modern remote sensing image change detection (CD), convolution Neural Network (CNN), especially U-shaped structure (UNet), has achieved great success due to powerful discriminative ability. However, UNet-based CNN networks usually have limitations in modeling global dependencies due to intrinsic locality of convolution operations. Transformer has recently emerged as an alternative architecture for dense prediction tasks due to global self-attention mechanism. However, due to limitation of hardware resources, pure Transformer methods generally lack the ability to capture global information at a low level. Based on these existing problems, we propose STransUNet, which combines Transformer and UNet architecture. STransUNet can not only capture shallow detail features at an early stage, but also model global context in high-level feature. In addition, we design an efficient feature fusion module named Cross-Enhanced Adaptive Fusion(CEAF). Our model mainly consists of three parts: encoder, fusion module and decoder. The decoder is a CNN-Transformer hybrid structure. CNN extracts multi-level feature information. Transformer encodes tokenized sequence to capture global context. CEAF module cross-enhances and adaptively fuses bi-temporal features to enhance feature representation. In decoding stage, we introduce a Cascaded Upsampling decoder(CUP). CUP progressively aggregates low-level CNN features and high-level Transformer features to full resolution. On four public CD datasets, our STransUNet achieves better CD results than six state-of-the-art algorithms.

We propose simple and effective models for the image annotation that make use of Convolutional Neural Network (CNN) features extracted from an image and word embedding vectors to represent their associated tags. Our first set of models is based on the Canonical Correlation Analysis (CCA) framework that helps in modeling both views - visual features (CNN feature) and textual features (word embedding vectors) of the data. Results on all three variants of the CCA models, namely linear CCA, kernel CCA and CCA with k-nearest neighbor (CCA-KNN) clustering, are reported. The best results are obtained using CCA-KNN which outperforms previous results on the Corel-5k and the ESP-Game datasets and achieves comparable results on the IAPRTC-12 dataset. In our experiments we evaluate CNN features in the existing models which bring out the advantages of it over dozens of handcrafted features. We also demonstrate that word embedding vectors perform better than binary vectors as a representation of the tags associated with an image. In addition we compare the CCA model to a simple CNN based linear regression model, which allows the CNN layers to be trained using back-propagation.

<p>Most of the currently known methods treat person re-identification task as classification problem and used commonly neural networks. However, these methods used only high-level convolutional feature or to express the feature representation of pedestrians. Moreover, the current data sets for person reidentification is relatively small. Under the limitation of the number of training set, deep convolutional networks are difficult to train adequately. Therefore, it is very worthwhile to introduce auxiliary data sets to help training. In order to solve this problem, this paper propose a novel method of deep transfer learning, and combines the comparison model with the classification model and multi-level fusion of the convolution features on the basis of transfer learning. In a multi-layers convolutional network, the characteristics of each layer of network are the dimensionality reduction of the previous layer of results, but the information of multi-level features is not only inclusive, but also has certain complementarity. We can using the information gap of different layers of convolutional neural networks to extract a better feature expression. Finally, the algorithm proposed in this paper is fully tested on four data sets (VIPeR, CUHK01, GRID and PRID450S). The obtained re-identification results prove the effectiveness of the algorithm.</p>

Deep learning has achieved great success in the field of computer vision, and the precision in image classification and image detection has surpassed humans. Therefore, this paper combines deep learning and medical image segmentation, focusing on how to improve the accuracy and speed of segmentation algorithm of medical exercise rehabilitation image. Aiming at the shortcomings of traditional medical image recognition methods, a medical exercise rehabilitation image segmentation algorithm based on hierarchical features of convolutional neural networks is proposed, this paper calls it as hierarchical features of convolutional neural networks (HFCNN). The algorithm firstly samples the convolution output of multiple layers in the convolutional neural network to a unified scale and combines them to construct a hierarchical feature. This hierarchical feature combines the structural information of objects contained in the shallow layer of the network with the semantic information of objects contained in the deep layers of the network, so it has a strong ability to express. Secondly, the image can be segmented into multiple super pixels by the super pixel segmentation algorithm. The classifier is trained using the hierarchical features of the super pixel, and then the classification result is mapped back to the pixel. Finally, a fully connected conditional random field algorithm including one-potential potential energy and paired potential energy is constructed. The corresponding energy function is used to smooth the classification result of the pixel, and the regional consistency and continuity of the pixel mark are improved. Compared with many classical convolutional neural network algorithms, this algorithm not only accelerates the network convergence speed, shortens the training time, but also significantly improves the accuracy of segmentation algorithm of medical exercise rehabilitation image, showing good practical value.

Scene text detection is a crucial step in end-to-end scene text recognition, a greatly challenging problem in computer vision. This paper proposes a novel scene text detection method that involves superpixel-based stroke feature transform (SSFT) and deep learning based region classification (DLRC). The SSFT is developed for candidate character region (CCR) extraction, which consists in partitioning an input image into several regions via superpixel-based clustering, removing most regions based on predefined criteria satisfied by the characters, and refining the remaining regions to obtain CCRs by computing a stroke width map. The character regions are identified from the CCRs using DLRC, in which several hand-crafted low-level features, i.e., color, texture, and geometric features, and some deep convolution neural network (CNN) based high-level features are first extracted from the regions, and then these features are fused by using two fully connected networks (FCNs) for region classification. In the DLRC step, the deep feature extraction CNN and the feature fusion FCNs are jointly trained. Next, the extracted character regions are merged to form candidate text regions, from which the final scene texts are detected. The proposed method is evaluated on three publicly available datasets: ICDAR2011, ICDAR2013, and street view text. It achieves F -measures of 0.876, 0.885, and 0.631, respectively, which demonstrate the effectiveness of the proposed scene text detection method.