Optimization of Image Similarity Comparison Algorithm Using Cosine Function

Abstract

With the explosive growth of digital image data, image similarity comparison, as the core technology of image retrieval and management, faces challenges such as low accuracy and poor robustness of traditional algorithms. To solve these problems, this paper proposes an optimized image similarity comparison algorithm based on the cosine function. First, the algorithm preprocesses the input image through adaptive median filter denoising, histogram equalization, and bilinear interpolation size normalization to eliminate interference factors. Then, it fuses the color features extracted by color moments (converted from RGB to HSV space) and texture features extracted by Local Binary Pattern (LBP) to form a comprehensive feature vector, and optimizes the feature vector through mutual information feature selection. Finally, the cosine function is used to calculate the similarity between feature vectors, and parallel computing and precomputation strategies are adopted to improve the calculation efficiency of cosine similarity. Experimental comparisons on Corel-1000 and Oxford Flowers 17 datasets show that the optimized algorithm (Cosine-Opt) has an F1-score of 89.9% and 85.6% on the two datasets respectively, which is significantly higher than traditional algorithms such as SAD-based and LBP-based. Robustness tests on distorted datasets (translation, rotation, scaling, Gaussian noise) show that the algorithm maintains high accuracy, and its computational efficiency is balanced with accuracy. This study provides an effective solution for efficient and accurate image similarity comparison.