According to data smooth theory, the best trade-off fractional order of 7.9, which the symmetric fractional B-spline wavelets can achieve between its smoothness and approximation texture, is deduced. A new thresholding formula which has self-adaptability with the order of the fractional spline wavelet, is derived, through improving the wavelet denosing model based on biva-Shrink neighboring coefficients relevance. Fractional spline wavelet transform is a new powerful tool for removing noise and irrelevant information from denoised images. Experimental result testifies the obtained theoretic result. Comparied with traditional denoising algorithms, it can achieve higher subjective and objective image qualities, especially for texture images. When the Barbara image with variance is no more than 10, the peak signal-to-noise ratio of the denoised image can reach 34.9842 and its geometric textures are well-protected after denoising by this algorithm.