An Efficient Scene Matching and Recognition Model for Large-Scale Identification Application

• Published in: Journal of Information Science and Engineering Vol. 41; no. 1; pp. 1 - 24
• Main Authors: Dai, Jun-Wei, Lu, Jian-Qiang, Mao, Yi-Min, Dai, Jing-Guo
• Format: Journal Article
• Language: English
• Published: Taipei 社團法人中華民國計算語言學學會 01.01.2025; Institute of Information Science, Academia Sinica
• Subjects: Adaptive algorithms; Algorithms; Clustering; Convolution; Datasets; Matching; Parallel processing; Parameters; Software; scene recognition; large-scale identification application; cross-domain; software development; parallel DCNN
• ISSN: 1016-2364
• DOI: 10.6688/JISE.202501_41(1).0001

Recently, software application techniques have evolved from the initial manual coding to efficient parallel computing and have been widely used in cross-domain disciplines. Scene matching and recognition model are one of the typical applications of software technology in cross-domain subjects. However, the model cannot efficiently handle large scene datasets. To address this problem, an efficient scene matching and recognition algorithm, ESMR-FFTCS, based on FFT convolution and Spark is proposed. First, a weight initialization strategy, WI-TDEK, based on image two-dimensional entropy and K-Means is presented to use the feature values extracted by clustering as the initial weights of the convolution kernel, solving the problem of slow model training convergence. Next, a parallel convolution operation strategy, PC-FFTS, is designed to improve the performance of convolution by parallel processing of denoised image data through Spark and FFT method. Finally, a load balancing strategy, LB-AWO, based on adaptive whale optimization algorithm, is proposed to obtain the node load value at parameters merging by iterative merit seeking, solving the problem of low efficiency in parameters parallel merging. The experimental results show that the recognition accuracy of EMST algorithm is suitable for fast recognition of large-scale scene datasets and has high training efficiency.