Computer Science and Technology is an international, peer-reviewed open access journal dedicated to advancing research the field of computer science and technology. The journal provides a rapid publication process to ensure wide dissemination of high-quality articles to scientists, professionals, and interested individuals worldwide. Our goal is to serve as an efficient, reliable, and trusted platform for scholars and readers, publishing cutting-edge research in the field.
Abstract: There have been many real-life applications utilizing deep learning, especially in the area of image classification. A common finding is that some domain data are highly skewed, which means that most of the information belongs to a small number of majority classes, and there is little or no information in the minority classes. Due to which in case of imbalanced data distribution, the majority of machine and deep learning algorithms are not effective or may fail when it is highly imbalanced. In this study, a comprehensive analysis of imbalanced dataset is conducted by considering deep learning-based well-known models. In particular, the best feature extractor model is identified and the current trend of latest feature extraction model is examined. Moreover, a bibliometric analysis is carried out from 1991 to 2022 in order to identify the global scientific research on the image classification of imbalanced mushroom dataset. In summary, our findings may offer researchers a quick benchmarking reference and alternative approach to assessing trends in imbalanced data distributions in image classification research.Abstract: There have been many real-life applications utilizing deep learning, especially in the area of image classification. A common finding is that some domain data are highly skewed, which means that most of the information belongs to a small number of majority classes, and there is little or no information in the minority classes. Due to which in case ...Learn More
Abstract: [Objective] With the release and application of ChatGPT, large model training technologies have started a new round of research boom, and this time is called the closest to AI AI. Through a comprehensive analysis of the current situation of the large model research based on NLP, the researchers can have a comprehensive understanding of the current research in this field in China. [Methods] Through the use of bibliometric and content analysis, analyzes the external characteristics and content of journal papers, through the annual post, journal cited, the core authors, core institutions, research keywords, research dimensions, China big model research has a more comprehensive cognition, but also summarizes some of the characteristics of the research in the field of China, also for the following research direction.[Results] China's research on large models has penetrated into many fields of society and is currently in the stage of vigorous development. [Conclusion] At present, the research in this field in China is in the second stage of the growth of scientific literature. With the continuous improvement of the basic theory of large model and the implementation of related applications, the research of large model based on NLP will enter a period of rapid development.Abstract: [Objective] With the release and application of ChatGPT, large model training technologies have started a new round of research boom, and this time is called the closest to AI AI. Through a comprehensive analysis of the current situation of the large model research based on NLP, the researchers can have a comprehensive understanding of the current ...Learn More
Abstract: Concrete compressive strength prediction is a key part of batching ratio design, the traditional concrete cube compressive strength standard test method is vulnerable to external influences, a random forest algorithm is proposed to predict the compressive strength of concrete, by optimizing the data of cement, blast furnace slag mixing, fly ash mixing, water content, high efficiency water reducer dosage, coarse aggregates, fine aggregates content, and age, etc., as raw material indicators. The optimized data set is used as the input data set, and the data set is divided to construct the model, and then the model parameters are optimized by defining the parameter network and performing the grid search cross-validation to construct the perfect models. The results show that the prediction accuracy of the prediction model based on the random forest algorithm (R²=0.91547) is much higher than that of the support vector regression model (R²=0.76802) and the decision tree model (R²=0.87539) and the error is small (RMSE=5.24087), which is of great significance for the research of compressive strength prediction model.Abstract: Concrete compressive strength prediction is a key part of batching ratio design, the traditional concrete cube compressive strength standard test method is vulnerable to external influences, a random forest algorithm is proposed to predict the compressive strength of concrete, by optimizing the data of cement, blast furnace slag mixing, fly ash mix...Learn More
Abstract: The data used for concrete compressive strength prediction is prone to produce outliers in the process of collection, which will have a certain impact on the accuracy of the model, so it is necessary to remove the abnormal data through data cleaning. In this paper, to address this problem, the random sampling consistency algorithm is introduced, and combined with the principal component analysis dimensionality reduction method, an improved anomalous data cleaning method is proposed. Firstly, the effectiveness of the data cleaning model of the random sampling consistency algorithm is verified, and the performance is compared with the commonly used isolated forest algorithm and K-means algorithm on the public dataset, and the performance indexes of the random sampling consistency algorithm model are obviously leading, and furthermore, the data are downgraded by using the principal component analysis and the selection rules of the optimization of the sampling points, so as to form the principal component analysis-random sampling consistency algorithm data cleaning model. Experiments show that the performance index of principal component analysis-random sampling consistency algorithm has no significant decrease compared with random sampling consistency algorithm, but the number of iterations is reduced to 492, and the time consumed is reduced to 196ms, which is 4.45% and 5.14% of the traditional random sampling consistency algorithm, respectively. The proposed improved method alleviates the volatility of the random sampling consistency algorithm, greatly reduces the number of iterations and time-consuming of the algorithm, and has a significant cleaning effect.Abstract: The data used for concrete compressive strength prediction is prone to produce outliers in the process of collection, which will have a certain impact on the accuracy of the model, so it is necessary to remove the abnormal data through data cleaning. In this paper, to address this problem, the random sampling consistency algorithm is introduced, an...Learn More
Abstract: The identification and classification of surface micro-patches in desertified grassland are crucial for dynamically monitoring grassland degradation. However, current traditional monitoring methods have several limitations such as high cost, difficult classification, and a small applicable area. In addition, the low resolution of satellite remote sensing makes it difficult to meet the precision requirements for dynamic monitoring of grassland desertification. In this study, we propose a ratio-adjusted vegetation index incorporating the spectral characteristics of surface micro-patches in desert steppe and the features of existing vegetation index models. The findings reveal an impressive overall recognition accuracy of 99.6% and a Kappa coefficient of 0.994. Notably, the proposed ratio-adjusted vegetation index outperforms four other vegetation index models, namely DNVI, SAVI, RVI, and DVI, by improving the overall accuracy by 39.5%, 37.8%, 39.3%, and 27.3%, respectively, with corresponding Kappa coefficients of 0.49, 0.472, 0.489, and 0.356. Additionally, the detection thresholds for rat hole, shadow, bare soil, and vegetation are determined to be (1, 1.0112), (1.0112, 1.0437), (1.0437, 1.1409), and (1.1409, ∞), respectively, while vegetation recognition accuracy remains consistently above 90%. Consequently, the RAVI method is found to be suitable for identifying, classifying, and extracting surface micro-patches of grassland desertification. Moreover, it can support the identification of various ground objects with the same vegetation index model, thereby providing a theoretical framework for classifying surface micro-patches in the desert steppe. Importantly, it opens up possibilities for monitoring the degradation of vast grassland areas.Abstract: The identification and classification of surface micro-patches in desertified grassland are crucial for dynamically monitoring grassland degradation. However, current traditional monitoring methods have several limitations such as high cost, difficult classification, and a small applicable area. In addition, the low resolution of satellite remote s...Learn More
Abstract: Wheat is a global food crop. Deep learning-based wheat head detection algorithm is helpful to simplify the planting process, reduce the planting cost and improve the wheat yield. However, due to the diversity of wheat traits and in-consistent growth cycles around the world, it is very important to build a detection model that can maintain high robustness and high accuracy in various scenarios. A deep learning-based wheat head detection algorithm was proposed. For the situation that domain data was scarce for wheat head detection, a semi-supervised pseudo label generation algorithm was proposed to semi supervised label the unlabeled data and use it for model training to improve the generalization performance in the new feature domain. In addition, in order to balance the speed and accuracy of wheat head detection algorithm, a multi model fusion mechanism was designed to fuse the detection algorithms with different characteristics. Finally, the open-source wheat head detection dataset was introduced and was used for algorithm verification. The algorithm obtained an average accuracy of 95.2% on its validation set and 76.1% on the test set, which showed that the proposed algorithm had robustness to cross domain scenes and excellent detection performance.Abstract: Wheat is a global food crop. Deep learning-based wheat head detection algorithm is helpful to simplify the planting process, reduce the planting cost and improve the wheat yield. However, due to the diversity of wheat traits and in-consistent growth cycles around the world, it is very important to build a detection model that can maintain high robu...Learn More
Abstract: The scheme of CP-ABE is widely used in data security protection in cloud outsourcing service. The architecture based on CP-ABE involves user, SP, PKG and CSP. Under this situation, PKG is trusted by default. However, in the real situation, except the default situation, there exists another scenario, where there is no SP and user communicates with PKG and CSP directly. In the above environment PKG may be untrusted. Once PKG is colluded with CSP or PKG is compromised, the adversary can get the decryption keys from PKG easily and the data preserved on CSP will be decrypted and leaked. In this paper, we first innovatively propose a scheme that can resist the attack from the aforementioned situations. The idea of the scheme is that a factor generated by the user is added to the secret key and the cipher text. The secret key is added the factor, which results that even PKG is compromised or untrusted, the secret key cannot be obtained. We apply the idea on the Li’s scheme and make it be suitable for the cloud outsourcing environment with an untrusted PKG. Finally, we prove the security of our scheme.Abstract: The scheme of CP-ABE is widely used in data security protection in cloud outsourcing service. The architecture based on CP-ABE involves user, SP, PKG and CSP. Under this situation, PKG is trusted by default. However, in the real situation, except the default situation, there exists another scenario, where there is no SP and user communicates with P...Learn More
Abstract: In the current big data environment, the scale of electronic archive data is increasing, due to its dispersion in different regions and systems, its dispersion has led to a sharp increase in the time cost of cross-domain access, and the misjudgment rate continues to rise under high concurrency conditions, which directly affects the accuracy and efficiency of electronic archive information retrieval. In order to solve the above problems, a cross-domain distributed storage electronic archives access technology model based on counting Bloom filter was proposed to support the accurate and efficient retrieval of cross-domain access to electronic archives. Compared with the traditional Bloom filter, the countable bloom filter significantly reduces the memory space consumption under the premise of ensuring the time complexity, and its key feature is that it can not only check the existence of elements, but also record the number of visits to the elements, and support dynamic addition and deletion of retrieved elements Reduce the false positive rate and improve the accuracy of the query. In order to solve the needs of large-scale data processing, the distributed architecture design is adopted to achieve the goal of decentralized server pressure and improve the overall response speed of the system. At the same time, the introduction of Zookeeper ensures the high availability of the system and ensures the cooperation between all nodes. After testing and practical application, the model can improve the efficiency of the file system, ensure the high availability of the system, and provide users with reliable and efficient electronic file management services.Abstract: In the current big data environment, the scale of electronic archive data is increasing, due to its dispersion in different regions and systems, its dispersion has led to a sharp increase in the time cost of cross-domain access, and the misjudgment rate continues to rise under high concurrency conditions, which directly affects the accuracy and eff...Learn More
Abstract: Traditional drug development requires a long process. Accurate prediction of drug-target binding affinity (DTA) by computer can greatly accelerate the drug development process. The key to predicting DTA is how to accurately mine the potential features of drugs and targets. To solve this problem, this paper proposes a DTA prediction model based on a multi-layer structure fused with bidirectional target features (MBDTA). DTA is predicted by this model through exploiting the representation of deep features of the drug-target pairs. MBDTA is split into three steps: Firstly, the initial features are obtained by encoding the drug molecules and targets through label encoding; Secondly, the initial features are fed into the graph neural network module and the recurrent neural network module to learn the potential features in them, respectively; Finally, the two sets of potential features are integrated and then the DTA is predicted exploiting a fully connected layer. Experimental results on the KIBA dataset show that MBDTA improves performance by an average of 32.42%, 3.84%, and 23.64% on the three metrics, MSE, CI, and r2 m , compared to the current state-of-the-art DTA prediction model.Abstract: Traditional drug development requires a long process. Accurate prediction of drug-target binding affinity (DTA) by computer can greatly accelerate the drug development process. The key to predicting DTA is how to accurately mine the potential features of drugs and targets. To solve this problem, this paper proposes a DTA prediction model based on a...Learn More
