Abstract: To address the issues of missed peeling and reliance on manual inspection for peeling rate assessment in mechanical lotus root peeling, this study proposes a method for identifying and targeted trimming of residual peel based on machine vision and an improved Pyramid Scene Parsing Network (PSPNet) for semantic segmentation. Given the complex curved surface of lotus root and the low contrast between peel and flesh, conventional mechanical peeling often leaves residual regions, which currently require manual trimming. The proposed method integrates visual recognition, precise localization, and automated trimming to overcome these limitations. First, an improved PSPNet semantic segmentation model is developed. Dilated convolutions with a dilation rate of 2 are incorporated into stage4 and stage5 of the backbone network ResNet50. This modification expands the receptive field without increasing the number of parameters, enabling the model to capture long-range contextual dependencies and better distinguish the subtle textural differences between lotus root peel and flesh. The original bilinear upsampler is replaced with a DySample dynamic upsampler, which uses 3×3 convolutions to generate spatially adaptive sampling weights based on the input feature map. This allows the decoder to reconstruct fine edge structures of residual peel more accurately, especially in irregularly shaped regions. Furthermore, a FreqFusion module is introduced into the decoder path. It adaptively fuses high-frequency edge details and low-frequency semantic information from multi-scale feature maps, significantly enhancing the model’s sensitivity to peel–flesh boundaries. Residual peel regions are localized using a global temporal image detection method. A fixed industrial camera captures the global video stream of the lotus root rotating 360° around the axis of the clamping mechanism. The trained improved PSPNet model processes the video stream frame by frame to generate binary segmentation masks of residual peel. Key frames are extracted when the Euclidean distance between the centroid of a detected residual region and the midline of the predefined region of interest (ROI) is minimized. This criterion ensures that the residual region is positioned directly facing the trimming cutter. Based on frame sequence analysis, the rotation angles corresponding to all residual regions are computed, and an optimized rotation path is planned. The stepper motor of the clamping mechanism then drives the lotus root to rotate incrementally, sequentially aligning each residual region with the trimming cutter. The trimming mechanism subsequently performs continuous feed motion to complete the targeted trimming operation with consistent cutting depth. An automatic calculation of the lotus root peeling rate is achieved using a pixel ratio method based on linear scan unfolded images. A line-scan camera acquires high-resolution images of the lotus root surface during a full rotation; these images are processed using the Hikvision industrial camera client software and distortion-corrected in conjunction with the captured front-view images. This yields an accurate circumferential unfolded image of the entire lotus root surface. The improved PSPNet model then detects residual peel regions in the unfolded image and generates a precise segmentation mask. Finally, the peeling rate is calculated by determining the pixel ratio of residual peel to the total lotus root surface area, providing a quantitative and objective cleanliness assessment. Experiments were conducted on a self-developed axial-feed lotus root peeling machine equipped with a trimming detection platform. The improved PSPNet model achieves a mean intersection over union (mIoU) of 92.15%, precision of 95.74%, and recall of 95.85%, representing improvements of 1.53, 0.92, and 0.75 percentage points, respectively, over the baseline PSPNet with ResNet50. The average angular positioning error of residual peel regions is as low as 4.6°, and the targeted trimming removal rate of these regions reaches 87.50%. As a result, the overall peeling rate is improved from 73.62% before trimming to 91.16% after trimming, an increase of 17.54 percentage points. These results demonstrate that the proposed method effectively eliminates residual peel with high precision and significantly enhances peeling quality. The research findings provide important guidance for the intelligent upgrading of lotus root peeling and processing equipment, and the proposed technical framework can be extended to other fruits and vegetables with similar peeling challenges.