Search

Numerous studies have shown deep hashing can facilitate large-scale image retrieval since it employs neural networks to learn feature representations and binary codes simultaneously. Despite supervised deep hashing has made great achievements under the guidance of label information, it is hardly applicable to a real-world image retrieval application because of its reliance on extensive human-annotated data. Furthermore, the pair-wise or triplet-wise unsupervised hashing can hardly achieve satisfactory performance due to the absence of local similarity of image pairs. To solve those problems, we propose a novel unsupervised deep hashing framework to learn compact binary codes, which takes the quadruplet forms as input units, called Unsupervised Deep Quadruplet Hashing with Isometric Quantization (UDQH-IQ). Specifically, by introducing the rotation invariance of images, the novel quadruplet-based loss is designed to explore the underlying semantic similarity of image pairs, which could preserve local similarity with its neighbors in Hamming space. To decrease the quantization errors, Hamming-isometric quantization is exploited to maximize the consistency of semantic similarity between binary-like embedding and corresponding binary codes. To alleviate redundancy in different bits, an orthogonality constraint is developed to decorrelate different bits in binary codes. Experimental results on three benchmark datasets indicate that our UDQH-IQ achieves promising performance.

State-of-the-art Re-ID methods based on unsupervised domain adaption transferred the domain knowledge by pre-training model on labeled source domain and fine-tuned the pre-trained model with pseudo labels generated by clustering samples on unlabeled target domain. Unfortunately, compared with supervised Re-ID methods, the performance of exiting clustering-based UDA Re-ID methods dropped sharply, since these methods utilized a not good enough baseline and seldom handled the noisy pseudo labels generated by clustering algorithm. In this paper, we try to address the UDA problem by designing a Strong Clustering-based Unsupervised Re-ID (SCURID) baseline for further research at first. Then, we integrate the hard labels and soft multilabels to learn more discriminative features on a unified angular regularized metric space. Specifically, we design the angular margin losses consisting of Hard Angular Margin Identification (HAMI) loss and Soft Angular Margin Identification (SAMI) loss. The HAMI loss can learn generalizable and discriminative features through hard pseudo labels generated by clustering on unlabeled target domain. The SAMI loss is proposed to refine the hard noisy pseudo labels through the soft multilabels obtained from peer network. Benefited from SCURID baseline and two angular margin losses, it enables the clustering-based UDA Re-ID model to alleviate the negative effect of noisy labels and toward more discriminability. Comprehensive and extensive experiments on three public available Re-ID datasets, e.g., Matket-1501, DukeMTMC-reID and MSMT17, demonstrate that our proposed method can outperform the state-of-the-art results on UDA Re-ID task.

With the advance of Convolutional Neural Network, deep hashing methods have shown the great promising performance in large-scale image retrieval. Without depending on extensive human-annotated data, unsupervised hashing is more applicable to image retrieval tasks compared to supervised methods. However, due to the lack of fine-grained supervised signals and multi-similarity constraints, most state-of-the-art unsupervised deep hashing algorithms cannot ensure the correct fine-grained similarity ranking for image pairs. In this paper, we propose a novel unsupervised deep multi-similarity hashing framework to learn compact binary codes by jointly exploiting global-aware and spatial-aware representations, called Unsupervised Deep Multi-Similarity Hashing with Semantic Structure (UDMSH). Specifically, to obtain distinguishing characteristics, we develop a sub-network by jointly learning global semantic structures from Convolutional Neural Network (CNN) and inherent spatial structures from Fully Convolutional Network (FCN). By computing the cosine distance for deep features from image pairs, we construct a similarity matrix with semantic structure, then utilize this matrix to guide hash code learning process. Based on it, we carefully design a multi-level pairwise loss to preserve the correct fine-grained similarity ranking. Furthermore, we introduce Hamming-isometric mapping into unsupervised hashing framework to decrease the quantization errors. Extensive experiments on three widely used benchmarks prove that our proposed UDMSH outperforms several state-of-the-art unsupervised hashing with respect to different evaluation metrics.

One of the major challenges of learning to hash in large-scale image retrieval is the projective transformation from raw image to binary space with preserving semantic similarity. Recently, several deep hashing methods show many excellent properties compared with traditional hashing based on hand-designed representation. However, most of the existing hashing models only pay attention to the semantic similarity between image pairs, ignoring the ranking information of retrieval results, which limits its performance. In this paper, a novel deep hashing framework, named Deep Top Similarity Hashing with Class-wise loss (DTSH-CW), is proposed to preserve semantic similarity between top images of ranking list and query images. In this proposed framework, CNNs architecture with batch normalization module is adopted to extract deep semantic characteristics. With integrating the position information of images, a top similarity loss is carefully designed to ensure the similarities between top images of ranking list and query images. Unlike pair-wise or triplet-wise loss, by directly leveraging the class labels, a cubic constraint based on Gaussian distribution is introduced to optimize objective function so as to maintain semantic variations of different classes. Furthermore, in order to solve discrete optimization problem, Two-Stage strategy is developed to provide efficient model training. Quantities of comparison experiments on three multi-label benchmark datasets show that our proposed DTSH-CW achieves promising performance compared to several state-of-the-art hashing methods.

Weather recognition is a significant technique for many potential computer vision applications in our daily lives. Generally, most existing works treat weather recognition as a single-label classification task, which cannot describe the weather conditions comprehensively due to the complex co-occurrence dependencies between different weather conditions. In this paper, we propose a novel Graph Convolution Networks with Attention (GCN-A) model for multi-label weather recognition. To our best knowledge, this is the first attempt to introduce GCN into weather recognition. Specifically, we employ GCN to capture weather co-occurrence dependencies via a directed graph. The graph is built over weather labels, where each node (weather label) is represented by word embeddings of a weather label. Furthermore, we design a re-weighted mechanism to build weather correlation matrix for information propagation among different nodes in GCN. In addition, we develop a channel-wise attention module to extract informative semantic features of weather for effective model training. Compared with the state-of-the-art methods, experiment results on two widely used benchmark datasets demonstrate that our proposed GCN-A model achieves promising performance.

Person re-identification (Re-ID) has attracted more attention in computer vision tasks recently and achieved high accuracy in some public available datasets in a supervised manner. The performance drops significantly when datasets are unlabeled, which limits the scalability of Re-ID algorithms in practical applications. Despite some unsupervised methods are proposed to address the scalability problem of Re-ID, it’s hard to learn discriminative feature representations due to the lack of pairwise labels in different camera views. To overcome this problem, we propose an end-to-end network named Adaptive Attention-Aware Network for unsupervised person re-identification. Specifically, we propose a novel adaptive attention-aware module that could be easily embedded into Re-ID architecture. The proposed module focuses on learning strong expressive relationship among channels of feature maps, and alleviating the key problems of Re-ID, e.g., occlusion and local deformation. In addition, we extract the camera-invariant features by adopting camera-style transfer feature learning since matching pairs in Re-ID suffers from appearance changes under different camera views. Besides, unsupervised hard negative mining is introduced to learn large intra-person appearance variance and discriminate high inter-person appearance similarity in an unlabeled target dataset with an auxiliary labeled dataset. Comprehensive experiments on three public available Re-ID datasets demonstrate that our method can achieve the state-of-the-art results of unsupervised Re-ID and is competitive with supervised learning.

For multi-label image retrieval based on deep hashing, the ultimate challenge is to map from the original image to binary space while preserving high-level semantic similarity. Recently, many supervised deep hashing approaches for multi-label image retrieval have been proposed to generate high-quality binary codes. However, most such methods are only introduced to learn simple similarity based on these image characteristics, therein ignoring complex multilevel semantic similarity with fine-grained features. In this paper, we propose a framework named deep hashing with fine-grained feature learning (DH-FFL) to preserve complex multilevel semantic similarity between multi-label image pairs. In this proposed model, compact bilinear pooling convolutional neural networks (CNNs) with normalization are adopted to extract fine-grained feature descriptors. In addition, a novel multilevel contrastive loss is designed to preserve multilevel semantic similarity by introducing a zero-loss parameter. Moreover, a multi-label classification loss is used to maintain the unique semantic structure of each image and maximize the distinguishing ability of binary codes. Comprehensive experiments on three benchmark datasets show that the proposed DH-FFL achieves promising performance compared with other state-of-the-art multi-label image retrieval applications.

Due to the advantages of retrieval speed and storage space, deep hashing methods have become a research hotspot in the field of large-scale image retrieval. Most of existing deep hashing methods pay close attention to similarity between images without images at the top of the ranking list similar to query targets. In the paper, a novel deep hashing model is proposed to preserve top images similar to the query images and optimize the quality of hash codes for image retrieval. Specifically, the optimized AlexNet is utilized to extract discriminative image representations and learn hashing functions simultaneously. The loss function based on acceleration strategy is designed to ensure similarity between returned images at the top of the ranking list and query images. In addition, we implement the model training in a batch-process fashion to low the image storage. Moreover, our extensive experiments on standard benchmarks demonstrate that our method outperforms several state-of-the-art deep hashing methods.

Multi-task learning aims to tackle various tasks with branched feature sharing architectures. Considering its diversity and complexity, discriminative feature representations need to be extracted for each individual task. Fixed geometric structures as a limitation of convolutional neural networks (CNNs) in building models, is also exists and poses a severe challenge in multi-task learning since the geometric variations will augment when we deal with multiple tasks. In this paper, we go beyond these limitations and propose a novel multi-task network by introducing the deformable convolution. Our design, the Deformable Multi-Task Network (DMTN), starts with a single shared network for constructing a shared feature pool. Then, we present task-specific deformable modules to extract discriminative features to be tailored for each task from the shared feature pool. The task-specific deformable modules utilize two new parts, deformable part and alignment part, to extract more discriminative task-specific features while greatly enhancing the transformation modeling capability. Experiments conducted on various multi-task learning types demonstrate the effectiveness of the proposed method. On multiple classification tasks, semantic segmentation and depth estimation tasks, our DMTN exceeds state-of-the-art approaches against strong baselines.

N-list and B-list have simply been proven to be highly effective for mining frequent itemsets. The main problem of the two novel structures is that they both need to encode each node of pre-order (or start order) and post-order (or finish order) code. This causes excessive memory consumption to mine frequent itemsets. In this paper, we propose SO-Sets based on SO-Tree, a more efficient data structure, to mine frequent itemsets. SO-Sets require only start-order (or finish-order) of each node, which makes it save lots of memory compared with N-list and B-list. Based on SO-Sets, we propose a new algorithm called FISO to mining frequent itemsets. To analyze the performance of algorithms, we conduct lots of experiments on five real datasets. Experimental results show that FISO algorithm has advantages in running time and size of main memory consumption.

As processed data is relatively dense or the support is small in weighted frequent patterns mining process, the number of frequent patterns which meet the conditions will be exponential growth, and mining all frequent patterns will need too much computation. Hence, mining the maximal weighted frequent patterns containing all frequent patterns has less calculation, and it has more utility value. Aiming at the process of maximal weighted frequent patterns mining, an efficient algorithm, based on WIdT-Trees, is proposed to discover maximal weighted frequent patterns. In the algorithm, WIdT-Tree is optimized from WIT-Tree. The dTidset strategy is used to calculate the weighted support of frequent k-itemsets, and the nodes with equal extended weighted support are pruned off in order to reduce the computational cost and decrease the search space complexity. Algorithms are tested and compared on real and synthetic datasets and experimental results show that our algorithm is more efficient and scalable.