Deep convolutional neural networks are increasingly used in various parallel embedded platforms such as mobile GPUs, AMD APUs, and FPGAs. At the same time, many new models have been developed for embedded platforms, such as MobileNet. In order to balance accuracy, speed and resource requirements and achieve cross-platform versatility, we have developed a software framework for in-depth research. Generated an OpenCL code that takes full advantage of parallel resources and improves the parallel efficiency of OpenCL code. Another advantage is that it optimizes and consolidates the network and compiles offline, making the entire application most efficient. MobileNets uses nonlongitudinal separable convolution (deep separable convolution) instead of standard convolution. Experiments with MobileNet have shown that the OpenCL code generation framework can significantly improve the efficiency of use.
Aircraft is a kind of valuable military equipment and transportation, so using target detection technology to detect ground aircraft in the optical remote sensing image has important research and application value. Although some achievements have been made in the relevant research, how to realize fast and effective ground aircraft target detection is still a challenging task because of the complex background of remote sensing image, large scale change and small imaging size, etc. Aiming at the application scenarios of multi-frame imaging, such as embedded detection and tracking system, this thesis proposes an aircraft target detection scheme based on hierarchical screening, which can improve the detection speed and reduce false alarm. Firstly, by analyzing the background characteristics, a target candidate region extraction method based on gray variance is adopted, and the acceleration is realized by integrating graph and shared computation. Then, the haar-like features are extracted in the candidate regions, which are then classified by the cascade AdaBoost classifier. Afterwards, a union-find-sets algorithm is used to merge the redundancy detection results and evaluate the confidence. Finally, the inter-frame correlation information is used to remove the false alarm. And we carried out experimental verification and proved the effectiveness of the algorithm.
In this paper, heterogeneous features extraction is conducted by deep learning for drug-related webpages classification. First, body text and image-label text are extracted through HTML parsing, and effective images are chosen by the FOCARSS algorithm. Second, text-based BOW model is used to generate text representation, and image-based BOW model is used to generate images representation. Webpages representation is generated by concatenating representations of text and images. Heterogeneous feature extraction are conducted by deep learning and classical methods, such as PCA, respectively. Feature selection is also conducted using information theory. Last, extracted features and selected features are classified. Experimental results demonstrate that the classification accuracy of features extracted by deep learning is higher than those of features extracted or selected by classical methods, and also higher than the accuracy of single modal classification.
In this paper, multi-modal local decision fusion is used for drug-related webpages classification. First, meaningful text are extracted through HTML parsing, and effective images are chosen by the FOCARSS algorithm. Second, six SVM classifiers are trained for six kinds of drug-taking instruments, which are represented by PHOG. One SVM classifier is trained for the cannabis, which is represented by the mid-feature of BOW model. For each instance in a webpage, seven SVMs give seven labels for its image, and other seven labels are given by searching the names of drug-taking instruments and cannabis in its related text. Concatenating seven labels of image and seven labels of text, the representation of those instances in webpages are generated. Last, Multi-Instance Learning is used to classify those drugrelated webpages. Experimental results demonstrate that the classification accuracy of multi-instance learning with multi-modal local decision fusion is much higher than those of single-modal classification.