Overcoming impurity interference: Cotton fiber length measurement based on FI-Unet combined with CAGAN image segmentation and restoration

Cotton fiber length is not only a key indicator for determining cotton quality, but also an important basis for parameter control and grading during cotton processing. To address the challenges of accurately detecting fiber length of machine-picked cotton with high impurity content during pre-purchase inspections, a novel method combining FI-Unet for image segmentation and CAGAN for image restoration is proposed, which effectively reduces impurity interference. Models for predicting upper half mean length (UHML) and length uniformity (UI) are developed by extracting grayscale, pixel count, and contour features from transmission images of fibers and applying ensemble learning. The prediction results demonstrate strong performance of the model. The UHML prediction achieves an R 2 (coefficient of determination) of 0.977, a RMSE (root-mean-squared error) of 0.274 mm, and a MAE (mean absolute error) of 0.210 mm, while the UI prediction yields an R 2 of 0.811, a RMSE of 1.054 %, and a MAE of 0.998 %. For impurity segmentation, the FI-Unet model incorporates residual blocks and a multi-scale spatial attention mechanism during downsampling to enhance low-level feature extraction and region focus. During upsampling, an improved CBAM (convolutional block attention module) module establishes a multi-scale channel-space attention mechanism to refine detail perception. The model achieves an accuracy of 99.87 %, an F1-score of 97.59 %, and an MIOU of 97.60 % in image segmentation. Subsequent CAGAN-based restoration of impurity-occlued regions attains a PSNR (peak signal-to-noise ratio) of 39.16 and a SSIM (structural similarity index) of 0.97. Post-restoration analysis shows further improvements in UHML and UI prediction, with R 2 increasing by 0.017 and 0.059, RMSE decreasing by 0.097 mm and 0.153 %, and MAE decreasing by 0.044 mm and 0.161 %, respectively. This study results contribute greatly to the enhancement of the accuracy and robustness of cotton fiber length detection under varying impurity conditions.