Convolutional neural networks (CNNs) have shown their power on many computer vision tasks. However, there are still some limitations, including their sensitivity to weight initialization and dependency to large scale labeled data. In this paper, we try to address these two problems by proposing a simple yet powerful CNN based denoising auto-encoder network which can be trained end-to-end in an unsupervised manner. The network architecture we use is a fully convolutional auto-encoder with symmetric encoder-decoder connections. The proposed method can not only reconstruct clean images from corrupted ones, but also learn abstract image representation through the reconstruction training. The encoder part of the network can be a good all-convolution network for classification. With the help of unsupervised pre-training, it achieves very competitive results even without extra unlabeled data. Further more, we show experimentally that our network also performs well in semi -supervised learning tasks.

Existing hashing methods normally define certain specific forms of hash functions, after which an objective function can be formulated to optimize the loss on training set to learn the parameters. However, in this way, the hash function will be tightly coupled with the generated objective in most cases. Moreover, since the objectives are generally formulated with binary quantization, most of them are nonconvex, which makes the optimization difficult and consequently decreases the similarity preserving performance of hashing. To solve this problem, we propose a novel pairwise correlation preserving framework to learn compact binary codes for hashing. First, we project each data into a metric space and represent it as a vector encoding the underlying local and global structure by pairwise correlation learning. Afterwards, pairwise correlation reconstruction (PCR), is further proposed to preserve the correlations of data between the metric space and the hamming space to learn binary codes. The PCR model is convex. Moreover, no specific hash functions are needed to be predefined and the steps of correlation learning and reconstruction are independent. The above characteristics make the optimization of PCR easily and efficiently, and thus leads to better preservation of data similarity in hamming space.

Text detection is a difficult task due to the significant diversity of the texts appearing in natural scene images. In this paper, we propose a novel text descriptor, SPP-net, extracted by equipping the Convolutional Neural Network (CNN) with spatial pyramid pooling. We first compute the feature maps from the original text lines without any cropping or warping, and then generate the fixed-size representations for text discrimination. Experimental results on the latest ICDAR 2011 and 2013 datasets have proven that the proposed descriptor outperforms the state-of-the-art methods by a noticeable margin on F-measure with its merit of incorporating multi-scale text information and its flexibility of describing text regions with different sizes and shapes.

In this paper, we propose a very deep fully convolutional encoding-decoding framework for image restoration such as denoising and super-resolution. The network is composed of multiple layers of convolution and deconvolution operators, learning end-to-end mappings from corrupted images to the original ones. The convolutional layers act as the feature extractor, which capture the abstraction of image contents while eliminating noises/corruptions. Deconvolutional layers are then used to recover the image details. We propose to symmetrically link convolutional and deconvolutional layers with skip-layer connections, with which the training converges much faster and attains a higher-quality local optimum. First, the skip connections allow the signal to be back-propagated to bottom layers directly, and thus tackles the problem of gradient vanishing, making training deep networks easier and achieving restoration performance gains consequently. Second, these skip connections pass image details from convolutional layers to deconvolutional layers, which is beneficial in recovering the original image. Significantly, with the large capacity, we can handle different levels of noises using a single model. Experimental results show that our network achieves better performance than recent state-of-the-art methods.

Part-based approaches, such as Deformable Part Model (DPM), have been actively studied for pedestrian detection in recent years. However, the overall pedestrian detection performance heavily depends on the detection accuracy of a single part detector, which is not so discriminative that usually easily affected by occlusions and deformations. To address this problem, in this paper, we propose a novel method enhancing the pedestrian detection performance by capturing the dependencies between different part detectors. Our attempt is to refine the overall detection accuracy by exploiting the joint state distribution of relevant part detectors. Firstly, we estimate the parameters for the joint state distribution, which expresses the dependencies between different part detectors, on the training set. Afterwards, a linear classifier is trained on the scores of the part detectors to assign them with different weights. At last, given a testing image, the joint state is inferred and the scores of part detectors are integrated to output the refined score. The experimental results carried out on the commonly used benchmark datasets show that the proposed method outperforms other state-of-the-art methods on pedestrian detection accuracy.

Due to the explosive growth of visual data and the raised urgent needs for more efficient nearest neighbor search methods, hashing methods have been widely studied in recent years. However, parameter optimization of the hash function in most available approaches is tightly coupled with the form of the function itself, which makes the optimization difficult and consequently affects the similarity preserving performance of hashing. To address this issue, we propose a novel pairwise correlation reconstruction framework for learning compact binary codes flexibly. Firstly, each data point is projected into a metric space and represented as a vector encoding the underlying local and global structure of the input space. The similarities of the data are then measured by the pairwise correlations of the learned vectors, which are represented as Euclidean distances. Afterwards, in order to preserve the similarities maximally, the optimal binary codes are learned by reconstructing the pairwise correlations. Experimental results are provided and analyzed on four commonly used benchmark datasets to demonstrate that the proposed method achieves the best nearest neighbor search performance comparing with the state-of-the-art methods.

Binary code learning techniques have been actively studied for hashing based nearest neighbor search in recent years. However, most existing techniques directly map the data into a Hamming space, which ignores the inherent property that original features may lie in different subspaces. To address this issue, this paper proposes a novel method by learning binary codes on the latent components decomposed from the original features of the data. We assume that each latent component represents an underlying feature subspace and similar/dissimilar data may contain the same/different latent components. As a result, the decomposition step distinguishes the original data better and consequently improves the discriminative power of binary codes. The experimental results of image retrieval tasks carried out on the commonly used benchmark datasets show that the proposed method outperforms other state-of-the-art methods on generating similarity-preserving binary codes.