At present, the same or similar style clothing retrieval is mainly text-based or content-based. The text-based algorithms often require massive labled samples, and the shortages of exist label missing and annotation difference caused by artificial subjectivity. The content-based algorithms usually extract image features, such as color, shape, texture, and then measured the similarity, but it is difficult to deal with background color interference, and clothing deformation due to different angles, attitude, etc. Aiming at these problems, clothing retrieval based on landmarks was proposed. The proposed method used cascaded deep convolutional neural network to locate the key points and combined the low-level visual information of the key point region as well as the high-level semantic information of the whole image. Compared with traditional methods, the proposed method can effectively deal with the distortion of clothing and complex background interference due to angle of view and attitude. Meanwhile, it does not need huge labeled samples, and is robust to background and deformation. Experiments on two large scale datasets Fashion Landmark and BDAT-Clothes show that the proposed algorithm can effectively improve the precision and recall.

Since the rapidly increasing data scale imposes a great computational challenge to the problem of Sparse Subspace Clustering (SSC), the existing optimization algorithms e.g. ADMM (Alternating Direction Method of Multipliers) for SSC are implemented in a sequential way which is unable to make use of multi-core processors to improve computational efficiency. To address this issue, a parallel SSC based on coordinate descent was proposed,inspired by a simple observation that the SSC can be formulated as a sequence of sample based sparse self-expression sub-problems. The proposed algorithm solves individual sub-problems by using a coordinate descent algorithm with fewer parameters and fast convergence. Based on the fact that the self-expression sub-problems are independent, a strategy was adopted to solve these sub-problems simultaneously on different processor cores, which brings the benefits of low computer resource consumption and fast running speed, it means that that the proposed algorithm is suitable for large scale clustering. Experiments on simulated data and Hopkins-155 motion segmentation dataset demonstrate that the proposed parallel SSC method on multi-core processors significantly improves the computational efficiency and ensures the accuracy when compared with ADMM.

Due to the disadvantages such as inefficiency of getting high-quality samples, bad impact of bootstrap to the whole learning-efficiency and final classifier performance in the negative samples bootstrap process of matrix-structural learning of cascade classifier algorithm. This paper proposed a fast learning algorithm-matrix-structural fast learning of cascaded classifier for negative sample inheritance. The negative sample bootstrap process of this algorithm combined sample inheritance and gradation bootstrap, which inherited helpful samples from the negative sample set used by last training stage firstly, and then got insufficient part of sample set from the negative image set. Sample inheritance reduced the bootstrap range of useful samples, which accelerated bootstrap. And sample pre-screening, during bootstrap process, increased sample complexity and promoted final classifier performance. The experiment results show that the proposed algorithm saves 20h in training time and improves 1 percentage point in detection performance, compared with matrix-structural learning of cascaded classifier algorithm. Besides, compared with other 17 human detection algorithms, the proposed algorithm achieves good performance too. The proposed algorithm gets great improvement in training efficiency and detection performance compared with matrix-structural learning of cascaded classifier algorithm.

The classic super-resolution algorithm via sparse coding has high computational cost during the reconstruction phase. In view of the disadvantages, a predictive sparse coding-based single image super-resolution method was proposed. In the training phase, the proposed method imposed a code prediction error term to the traditional sparse coding error function, and used an alternating minimization procedure to minimize the resultant objective function. In the testing phase, the reconstruction coefficient could be estimated by simply multiplying the low-dimensional image patch with the low-dimensional dictionary, without any need to solve sparse regression problems. The experimental results demonstrate that, compared with the classic single image super-resolution algorithm via sparse coding, the proposed method is able to significantly reduce the reconstruction time while maintaining super-resolution visual effect.