Purpose <p>Acquiring accurate in-situ sediment particle size data in turbid aquatic environments remains a challenge for optical techniques, due to pervasive issues like color distortion, blurring, and low contrast. To address this, we established a sediment particle image dataset under controlled turbid bentonite conditions (NTU = 5, 10, 15) and magnifications (4×, 10×, 20×) and developed a multi-scale fusion algorithm that adaptively integrates contrast enhancement, denoising, and deep learning to restore images.</p> Methods <p>For sediment images under varying turbidity (NTU = 5, 10, 15) and magnification (4×, 10×, 20×), our algorithm adaptively integrates the strengths of contrast enhancement, denoising, and deep learning algorithms, significantly suppressing color distortion and enhancing particle boundary clarity.</p> Results <p>The method significantly increased RGB histogram standard deviation (&gt; 72% overall), enhancing contrast and obscured detail recovery. Targeted compensation specifically enhanced the blue channel (B) by &gt; 100%, effectively counteracting bentonite-induced blue-light attenuation to restore natural color balance. Crucially, particle size identification using restored images yielded results significantly closer to sieve analysis than original images. Deviations in key sediment grading parameters—uniformity coefficient (<i>C</i>_<i>u</i>) and curvature coefficient (<i>C</i>_<i>c</i>)—decreased from 7.68% to 5.94% and 2.00% to 1.11%, respectively.</p> Conclusion <p>Our approach effectively overcomes optical limitations in turbid sediment-laden waters, offering an effective tool for improving in-situ particle size analysis. This advancement holds promise for applications in marine geotechnical engineering and environmental sediment monitoring where turbidity impedes traditional methods.</p>