The Research On Key Classification Technologies Of Color Image Region Of Interest Based On Genetic Algorithm And Neural Network

Image semantic classification is an important and challenging task in image retrieval research based on semantics. Traditional image classification techniques mainly classify the images according to the similarity of the image visual features. But the semantic content contained in image could not be accurately expressed by the underlying characteristics of the image, for the significant differences between the underlying characteristics of the image and human understanding, that is, there is "semantic gap" between the underlying visual characteristics of the image and semantics contained in image.This paper takes the color images as the data source and does an in-depth study on the key technologies and the main algorithms in image semantic classification. The work is primarily focused on:(1) An extraction algorithm based on image region of interest is proposed. The algorithm introduces the image complex analysis on the bases of traditional segmentation algorithm, that is, it would take different segmenting algorithm according to the difference of image's complexity. This solves the problems brought with exact image division. Image area extraction only expresses the image on its object-level and distinguishes the image content in the use of its local characteristics. The experiments prove that the algorithm could maintain or even improve the recognition rate, while reduce the computational complexity.(2) Through analyzing the characteristics of texture, it is found that, in the actual texture recognition, certain type of texture feature is more obvious, so the recognition need only a few parameters. But certain types of texture feature is not very obvious, it needs more characteristic parameters as input parameters to achieve a higher recognition rate to identify the image. Based on the analysis of traditional extraction method of texture features values, this paper adopts the Tamura texture feature extraction method and solves the problem of low extraction efficiency of values. Moreover, texture properties generally have intuitive visual sense, so this method is more in line with the human understanding to image.(3) For the learning convergence speed of neural network is too slow and it could not guarantee to converge to the global minimum point. The paper proposes an algorithm to improve the neural network by genetic algorithm. The range of cross-operating in the algorithm is wide, for the multiple individual components is randomly used, so it is easier to maintain the colony diversity. The selection operation of algorithm is with pertinence, this ensures that the colony could be stably constringed to the optimal solution. The experimental results show that the method is with high recognition rate and high convergence speed. (4) The paper analyzes the semantic mapping model, describes the related theories of semantic mapping and their methods, and designs neural network classifiers based on genetic algorithm to complete the mapping between the image low-level features and high-level semantic features which achieves the goal of automatic acquisition of image semantics. Finally, it combines with the extracted image texture features value and completes the image region classification of interest. The system is been compared with the SIMPLIcity system in Corel image library, and the experiment results show that the system has better retrieval precision.