In recent years, deep learning achieves very good results and has great improvement in object detection. However, in some special scenes, for example, when it is required to simultaneously detect objects with greatly different scales (difference greater than 100 times), common object recognition methods' performance will drop dramatically. Aiming at the problem of recognizing and locating objects with super-large-scale variance in the same scene, the You Only Look Once version3 (YOLOv3) framework was improved, the image pyramid technology was combined to extract the multi-scale features of the image. And in the training process, the strategy of using dynamic Intersection over Union (IoU) was proposed for different scale objects, which was able to better solve the problem of sample imbalance. Experimental results show that the proposed model significantly improves the recognition ability of super-large and super-small objects in the same scene. The proposed model has been applied to the airport environment and achieved good application results.

Aiming at the problems of inadequate semantic understanding, improper summary sentences and inaccurate summary in the field of Natural Language Processing (NLP) abstractive automatic summarization, a new automatic summary solution was proposed, including an improved word vector generation technique and an abstractive automatic summarization model. The improved word vector generation technology was based on the word vector generated by the skip-gram method. Combining with the characteristics of abstract, three word features including part of speech, word frequency and inverse text frequency were introduced, which effectively improved the understanding of words. The proposed Bi-MulRnn+ abstractive automatic summarization model was based on sequence-to-sequence (seq2seq) framework and self-encoder structure. By introducing attention mechanism, Gated Recurrent Unit (GRU) gate structure, Bi-directional Recurrent Neural Network (BiRnn) and Multi-layer Recurrent Neural Network (MultiRnn), the model improved the summary accuracy and sentence fluency of abstractive summarization. The experimental results of Large-Scale Chinese Short Text Summarization (LCSTS) dataset show that the proposed scheme can effectively solve the problem of abstractive summarization of short text, and has good performance in Rouge standard evaluation system, improving summary accuracy and sentence fluency.

Concerning low accuracy of age classification of face images under unrestricted conditions, a new method of face age classification under unconstrained conditions based on deep Residual Networks (ResNets) and large dataset pre-training was proposed. Firstly, the deep residual networks were used as the basis convolutional neural network model to deal with the problem of face age classification. Secondly, the deep residual networks were trained on the ImageNet dataset to learn the expression of basic image features. Thirdly, the large-scale face age images IMDB-WIKI was cleaned, and the IMDB-WIKI-8 dataset was established for fine-tuning the deep residual networks, and migration learning from the general object image to face age image was achieved to make the model adapt to the distribution of the age group and improve the network learning capability. Finally, the fine-tuned network model was trained and tested on the unconstrained Adience dataset, and the age classification accuracy was obtained by the cross-validation method. Through the comparison of 34/50/101/152-layer residual networks, it could be seen that the more layers of the network have the higher accuracy of age classification. And the best state-of-the-art age classification result on Adience dataset with the accuracy of 65.01% was achieved by using the 152-layer residual network. The experimental results show that the combination of deeper residual network and large dataset pretraining can effectively improve the accuracy of face age classification.

Due to the deficiency of ubiquitous charging stations (or stakes) and short driving distances of Electric Vehicle (EV), many people are hesitant to use EV. To reduce users' anxiety about limited battery capacity and lower fees due to frequent charging and making detour to charge, a matching theoretic Traveling Plan-aware Charging Scheduling (TPCS) scheme was proposed. Firstly, preference lists of EV users and charging stations were constructed respectively according to traveling plans of EV and their electricity demand at each charging station. Secondly, a many-to-one matching model was established between EV users and charging stations. Finally, interfaces of charging stations were allocated to optimize the system total utility. Compared with the Random Charging Scheduling (RCS) algorithm and Only utility of Electric Vehicle concerned Scheduling (OEVS) algorithm, the system total utility of TPCS was increased at most by 39.3% and 5% respectively. In addition, TPCS guaranteed the satisfactory ratio of EV users to be above 90% when charging demand of EV users was light, which is higher than that of RCS. The proposed algorithm can effectively improve the system total utility and satisfactory ratio of EV users, and reduce the computational complexity.

A method based on projection entropy for image recognition was introduced in this paper. Since original definition of projection entropy does not make full use of image information and is not scale invariant, a new definition was proposed. The Local Projection Entropy (LPE) of normalized image was used for image recognition. In the process of recognition, firstly, Gaussian Mixture Model (GMM) of training set images’ LPE was obtained by Expectation Maximization (EM) algorithm. Then the Mahalanobis distance of target image’s LPE and GMM was calculated. The category of image was determined according to the distance discriminant law. Computer vision laboratory databases of Columbia university were used in the experiments, and the results show that the proposed algorithm is an effective approach for image recognition and has a proper structure for parallel computing.

In the paper, on the basis of analyzing the structure of dynamic information database, and in combination with the feature of the power system, the key technologies of concurrency data processing, memory-mapped file, disk cache management mechanism and associated data storage were discussed, and the data sampling flow and hybrid compression algorithm were also introduced in detail. The application case in the automatic system of power grid dispatching was introduced and the result proves that the dynamic information database can meet the performance requirement of high-speed data processing.

Compiler is an important part of embedded system software, it has great impact on software development of embedded system. This paper analyzed the structure of an compiler automatic generation tool genmd, based on the idea of combining ADL with retargetable compilers, and put emphasis on abstracting and modeling of instruction recognition and machine description generation. Meanwhile, a method was provided and implemented to solve the limitation of genmd that was unable to support instructions of branch and jump.