Sperm morphology assessment plays a vital role in semen analysis and the diagnosis of male infertility. By quantitatively analyzing the morphological characteristics of the sperm head, midpiece, and tail, it provides essential insights for assisted reproductive technologies (ARTs), such as in vitro fertilization (IVF). However, traditional manual evaluation methods not only rely on staining procedures that can damage the cells but also suffer from strong subjectivity and inconsistent results, underscoring the urgent need for an automated, accurate, and non-invasive method for multi-sperm morphology assessment. To address the limitations of existing techniques, this study proposes a novel method that combines a multi-scale part parsing network with a measurement accuracy enhancement strategy for non-stained sperm morphology analysis. First, a multi-scale part parsing network integrating semantic segmentation and instance segmentation is introduced to achieve instance-level parsing of sperm, enabling precise measurement of morphological parameters for each individual sperm instance. Second, to eliminate measurement errors caused by the reduced resolution of non-stained sperm images, a measurement accuracy enhancement method based on statistical analysis and signal processing is designed. This method employs an interquartile range (IQR) method to exclude outliers, Gaussian filtering to smooth data, and robust correction techniques to extract the maximum morphological features of sperm. Experimental results demonstrate that the proposed multi-scale part parsing network achieves 59.3% APvolp, surpassing the state-of-the-art AIParsing by 9.20%. Compared to evaluations based solely on segmentation results, the integration of the measurement accuracy enhancement strategy significantly reduces measurement errors, with the largest reduction in errors for head, midpiece, and tail measurements reaching up to 35.0%.