Publications
Sort:
Issue
Automatic identification and targeted peeling of residual epidermis on lotus root based on an improved PSPNet
Transactions of the Chinese Society of Agricultural Engineering 2026, 42(14): 365-375
Published: 30 July 2026
Abstract PDF (2.7 MB) Collect
Downloads:0

Lotus root is the vegetable rhizome of the lotus plant in Asian areas. Mechanical peeling is one of the most important steps during harvesting. However, the residual regions can be left after conventional mechanical peeling and currently require manual trimming, due to the complex curved surface of the lotus root and the low contrast between peel and flesh. Furthermore, the peeling rate assessment has been limited to manual inspection. This study aims to identify the residual epidermis on the lotus roots for target peeling using machine vision. An improved Pyramid Scene Parsing Network (PSPNet) was proposed for semantic segmentation. Three procedures were constructed, including visual recognition, precise localization, and automatic trimming. 1) Visual recognition. Dilated convolutions with a dilation rate of two were incorporated into Stage 4 and Stage 5 of the backbone network ResNet50. The receptive field was expanded without increasing the number of parameters after modification. Long-range contextual dependencies were captured to better distinguish the subtle textural differences between lotus root peel and flesh. The original bilinear upsampler was replaced with a DySample dynamic upsampler. The 3×3 convolutions were used to generate spatially adaptive sampling weights, according to the input feature map. As such, the decoder was used to more accurately reconstruct the fine edge structures of residual peel, especially in irregularly shaped regions. A FreqFusion module was introduced into the decoder path. The sensitivity to peel–flesh boundaries was significantly enhanced to adaptively fuse the high-frequency edge features and low-frequency semantic information from multi-scale feature maps. 2) Precise localization. Residual peel regions were localized to detect the global temporal images. A fixed industrial camera was used to capture the global video stream of the lotus root via the rotation of 360° around the axis of the clamping mechanism. The video stream was processed frame by frame to generate binary segmentation masks of residual peel. Key frames were then extracted to minimize the Euclidean distance between the centroid of a detected residual region and the midline of the predefined region of interest (ROI). The residual region was positioned directly face-to-face with the trimming cutter. The rotation angles corresponding to all residual regions were computed using frame sequence analysis. An optimal rotation path was planned after calculation. The stepper motor of the clamping mechanism was used to incrementally drive the rotation of the lotus root, sequentially aligning each residual region with the trimming cutter. The trimming mechanism subsequently performed continuous feed motion to realize the target trimming with consistent cutting depth. 3) Automatic trimming: The peeling rate of lotus root was achieved using a pixel ratio with linear scan unfolded images. A line-scan camera was used to acquire the high-resolution images of the lotus root surface during full rotation; Hikvision industrial camera client software was used to correct the distortion with the front-view images. An accurate image was obtained from the unfolded circumference of the entire lotus root surface. A precise segmentation mask was then generated to detect the residual peel regions in the unfolded image. Finally, the peeling rate was calculated to determine the pixel ratio of the residual peel to the total lotus root surface area. Experiments were conducted on a self-developed axial-feed lotus root peeling machine with a trimming detection platform. The improved PSPNet model was achieved in a mean intersection over union (mIoU) of 92.15%, precision of 95.74%, and recall of 95.85%, which was improved by 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 was 4.6°, and the fixed-point targeted peeling removal rate of residual lotus root peel regions reached 87.50%. As a result, the overall peeling rate was improved from 73.62% to 91.16% after trimming, with an increase of 17.54 percentage points. The residual peel effectively enhanced the peeling quality with high precision. The research findings can provide important guidance for the intelligent upgrading of lotus root peeling equipment. The technical framework can be extended to the rest of the fruits and vegetables during peeling.

Issue
Regenerated buds tracking and regenerative ability evaluation of ratooning rice using Micro-CT imaging and improved DeepSORT
Transactions of the Chinese Society of Agricultural Engineering 2023, 39(11): 165-174
Published: 15 June 2023
Abstract PDF (3.3 MB) Collect
Downloads:0

Regenerative ability is closely related to the number of regenerated buds in the ratooning season, even the yield of ratooning rice. The traditional detection cannot fully meet the large-scale production in the rice regenerated buds, due to contact damage, subjective inefficiency, and low repeatability. In this study, a multi-target tracking of regenerated buds was proposed for high-precision counting using Micro-CT (computed tomography) and an improved DeepSORT. Micro-CT imaging was first adopted to capture the cross-section video stream of rice stem in the ratooning season. Then, the YOLOv5s network was used as the tracking detector of regenerated buds, while the improved DeepSORT tracking was studied to achieve the accurate tracking and counting of rice regenerated buds. The ID discrepancy was then optimized to improve DeepSORT tracking. The matching accuracy of the multi-target tracking increased using the feature of continuity between each frame of CT tomogram images, indicating the substantially improved ID switch. Finally, the height of the regenerated buds was calculated to discriminate the effectively regenerated buds using the location information of the tracking object. The experimental results showed that the Mean Average Precision (mAP) values of YOLOv5s were 97.3% and 99.1% for the regenerated buds and stalks, respectively, in the target detection. The multi-object tracking accuracy (FMOTA ) , higher order tracking accuracy(FHOTA) , and ID switch of the improved DeepSORT were 77.61%, 61.73%, and 6, respectively, in the multi-target tracking, compared with the original. Furthermore, the FMOTA and FHOTA were improved by 1.51% and 8.5%, respectively, whereas the ID switch was improved by 94%. The multi-object tracking efficiency of the DeepSORT and the improved were 25, and 24 frames per second, respectively, without a significant decrease in the efficiency. The system and manual measurements of 104 pots of ratooning rice were used to verify the regenerated buds, where the correlation coefficient square R2 of 0.983, the root mean square error of 3.460, and the average absolute percentage error of 5.647%, indicating a better consistency with the manual measurement. The ratio of the regenerated buds to the number of the stem was computed for the early ratooning ability of rice. The correlation analysis was also performed between the ratooning ability of two rice varieties in 38 pots and the actual yield in the ratooning season. It was found the R2 values were 0.795 and 0.764, respectively, indicating a significant positive correlation between the regenerative ability and rice yield. In conclusion, the novel nondestructive way was achieved to detect the regenerated buds in the early regenerative ability measurement. The finding can also provide important technical support for the ratooning rice breeding.

Issue
Rice panicle tracking and length extraction based on optical flow pretreatment and StrongSORT
Transactions of the Chinese Society of Agricultural Engineering 2025, 41(11): 146-155
Published: 15 June 2025
Abstract PDF (3 MB) Collect
Downloads:3

The panicle number and length are two of the most crucial indicators of rice yield. Accurate acquisition of the panicle traits is of great significance to rice breeding and genetic research. However, the traditional measurements of panicle traits cannot fully meet the large-scale production in recent years, due to the contact damage, subjective inefficiency, and low repeatability. Therefore, it is very urgent to develop the observation and identification for the accurate and efficient measurement of rice panicle traits. In this study, the rice panicle tracking and traits extraction were proposed using optical flow preprocessing and the StrongSORT algorithm. Initially, a series of experiments was conducted to capture the 200 rotating rice videos. The dataset was divided into the training and testing sets in a ratio of 8:2. Subsequently, the Gunnar Farneback optical flow algorithm was employed to preprocess the videos in order to reduce the occlusion. The Convolutional Block Attention Module (CBAM) attention mechanism was then integrated into the YOLOv8-seg network in order to enhance the target detection and segmentation of rice panicles. Finally, the StrongSORT algorithm was utilized to realize the multi-target tracking and the counting of rice panicles. The Zhang-Suen skeleton extraction was applied to determine the length of the rice panicle with the largest panicle after detection. Moreover, a motion prior model was constructed with the movement trajectories and velocities of the potted rice. The position of rice panicles was predicted in the next frame. The ID switches were reduced to prevent the panicle tracking failures and duplicate counting caused by occlusion. The results demonstrated that high accuracy of the tracking was achieved to detect the rice panicle. The mean average precision of the improved YOLOv8-seg model reached 81.1%, with an increase of 8.7 percentage points, compared with the original YOLOv8-seg model. Furthermore, the mAP of the YOLOv8-seg model was improved to 95.0% after optical flow preprocessing, indicating a substantial enhancement of 13.9 percentage points over the unprocessed model. In rice multi-target tracking, the combination of optical flow preprocessing, the improved YOLOv8-seg, and StrongSORT was achieved in a multi-target tracking accuracy of 85.58% and a high-order tracking accuracy of 64.06%, which were improved by 11.83 and 9.53 percentage points, respectively, compared with the combination without optical flow preprocessing. The number of ID switches was significantly reduced from 891 to 275, with a decrease of 69.2%. In terms of counting accuracy, the combination was achieved in a coefficient of determination (R2) of 0.969 6, a mean absolute percentage error of 2.15%, and a root-mean-square error of 1.87, compared with the actual values. In panicle length extraction, the R² value was 0.940 8, the MAPE was 4.07%, and the RMSE was 0.47. The combination of the optical flow preprocessing, improved YOLOv8-seg, and StrongSORT effectively reduced the interference among overlapping panicles and the ID switch. The accuracy and multi-target tracking were enhanced after detection at the same time. The finding can also provide the technical pathway for the rice panicles and length measurement in rice breeding.

Open Access Issue
Leaf grading for cotton verticillium wilt based on VFNet-Improved and Deep Sort
Journal of Intelligent Agricultural Mechanization 2023, 4(2): 12-21
Published: 15 May 2023
Abstract PDF (27.9 MB) Collect
Downloads:1

Cotton is one of the most important economic crops in the world, and verticillium wilt is the number one disease in major cotton production areas in the world. Verticillium wilt causes the leaves to wilt, fade and even fall off by infecting the roots of cotton, resulting in severe decline both cotton quality and its yield. The national standard divides the leaves suffering from verticillium wilt into five grades. The traditional detection method mainly relies on manual labor, which has problems such as subjectivity, inefficiency, and poor repeatability. The cotton verticillium wilt disease classification method, which is based on VFNet-Improved, Deep Sort, and collision line matching mechanisms as the main algorithm framework, realizes the statistics of the number of diseased leaves and the classification of the disease level under the condition of rotating video input. Based on the VFNet target detection network, the research first combined multi-scale training, dynamic convolution and other optimization methods to achieve accurate positioning of diseased leaves in rotating videos. Then the Deep Sort tracker was used to realize the correlation of the same leaf in the front and back frames, and a mask collision line matching mechanism was designed for the ID jump problem in the tracking process; finally, OpenCV was used to perform feature extraction and disease classification of leaves passing the mask line. The results showed that the VFNet-Improved model could achieve the best effect in the object detection algorithms with 0.906 mAP75, which was 0.012 higher than the VFNet model, and the FPS reached 12.9 frames/s. The tracking result MOTA of the Deep Sort was 0.835. R2, RMSE, MAE and MAPE of VFNet-Improved were 0.890, 5.138, 4.300 and 14.967% respectively, which performed high consistency with manual measurements. In conclusion, this study has demonstrated a novel tool for the accurate and efficient evaluation of cotton verticillium wilt, which would be of great significance for the cotton breeding and genetic analysis research.

Total 4