An improved lightweight YOLOv4 model was proposed to realize the accurate, real-time, and robust automatic sorting of bare and shelled shrimp in the shrimp mechanical shelling process under complex scenarios. The CSP-Darknet53 network was replaced by the GhostNet in the YOLOv4 structure. The ability of the model was then improved to extract the features adaptively. The calculation of model parameters was also simplified after improvement. The GhostNet network was used for the YOLOv4 backbone feature extraction, in order to reduce the network model complexity, and the model parameters for better storage capacity and detection efficiency. A lightweight attention mechanism was introduced into the Resblock module of the YOLOv4 backbone feature extraction network, in order to enhance the feature extraction capability of the backbone feature extraction network. The SE attention mechanism module was used to enhance the attention between feature channels. The attention of the network model was improved to the shrimp shell by fitting the relevant feature information to the target channel and suppressing invalid information. The model recognition accuracy was improved to reduce background interference. The original GIoU loss function was replaced with a CIoU loss function to improve the regression effect of the prediction frame. The CIoU loss function made the data obtained from non-maximal suppression more reasonable and efficient. Furthermore, the prediction frame was more accurate to minimize the distance between the centroids of the detection frame and the labelled frame. The lightweight GhostNet-YOLOv4 model was compared with the YOLOv7, EfficientNet Lite3-YOLOv4, ShuffleNetV2-YOLOv4, and MobilenetV3-YOLOv4 models. The results showed that the GhostNet-YOLOv4 model shared the lowest number of parameters and computational effort. An ablation comparison experiment was designed to verify that replacing the backbone feature extraction network and embedding the SE attention mechanism optimized for the module. The replacement of the CSP-Darknet53 backbone feature extraction network with the GhostNet resulted in a 2.9 percentage point improvement in the mAP and a significant reduction in the number of model parameters and output weights, compared with the original model. The addition of the SE attention mechanism improved the anti-interference and feature extraction ability, whereas the mAP was improved by 1.8 percentage points. After replacing the GIoU loss function with the CIoU one, the shrimp recognition accuracy was further improved, where the mAP was improved by 1.4 percentage points. According to the actual operating environment of the shrimp shell inspection test bed, two types of image datasets were produced, namely bare flesh shrimp and shelled shrimp. The GhostNet-YOLOv4, YOLOv3, YOLOv4, and MobilenetV3-YOLOv4 models were used for testing. The results show that the GhostNet-YOLOv4 model achieved detection accuracy and speed of 95.9% and 25 frames/s, respectively. The GhostNet-YOLOv4 model outperformed all other models in terms of detection speed under the condition of guaranteed detection accuracy. The performance of the GhostNet-YOLOv4 network model was evaluated to identify the shrimp shells for four treatments with the changes in light brightness, speed, shrimp posture, and shrimp species. The shrimp shell detection test showed that the average accuracy of shrimp shell recognition reached 90.4%, fully meeting the operational requirements. It indicates that the test bench was suitable for installation on mobile-embedded devices. The GhostNet-YOLOv4 network model still shared excellent generalization performance, when identifying other species of shrimp shells outside the sample set, with an average accuracy of 87.2%.

Weed damage is one of the main reasons leading to the reduction of rice yield and quality. The elastic tooth pulling weed control can not only effectively remove weeds between rows and plants, but also solve the problem of weed mechanical control in the near root zone of rice. At present, it has become the main technology of mechanical control of weed damage in rice field. However, the existing tine operation process does not distinguish between rice and grass growth areas, and uses constant operation parameters and the same structural parameters, which makes it difficult to take into account the weeding rate and seedling injury rate at the same time in the operation process, so it is urgent to further study the technique and equipment of tine seedling protection and weeding. The methods of protecting seedlings and weeding in elastic toothed paddy field are mainly classified as operation parameter optimization, structural parameter improvement and operation mode innovation. The optimization of operation parameters is currently the most widely used way of protecting seedlings and weeding with elastic teeth, but the optimal operation parameters obtained are only suitable for the current rice field environment, and the method of protecting seedlings and weeding is lack of universality. In terms of structural parameter improvement, take the elastic tooth weeding as an example. At present, the operation process of the elastic tooth swing type rice field weeder does not distinguish between seedling and grass growth areas. The improvement of the elastic tooth structure is restricted by the injured seedling rate, and the space for improving the weeding rate is limited. As for the innovation of operation mode, facing the complex environment of paddy field and the dense planting conditions of rice, there are still technical bottlenecks in the accurate division of rice grass growth area by conventional perception methods such as vision and laser, and the effective method of fast switching operation mode between seedling area and grass area needs to be further studied. In view of the special environment of rice field, this study proposed a method of dividing the seedling grass growth area based on tactile perception, and designed a weeding machine with variable comb and oblique tooth structure, adjustable operation depth and operation range parameters, and fast switching operation mode in the seedling grass growth area. A combination of flat pressure comb teeth and inclined swing teeth was adopted to design a rapid mode switching mechanism, and the values of the pressure amplitude and swing amplitude were determined through parameter optimization. A rice grass growing region perceptron was designed by using piezoelectric film sensor and bending sensor. Through data acquisition, filtering, segmentation and other processing, the rice seedling and weed growing regions were accurately divided. Based on the division results, the operation mode is automatically switched to the flat pressure comb operation mode with low seedling damage rate in the root zone of rice seedlings, and the pressure swing oblique tooth operation mode with high weed control rate in the weed growth area, so as to realize seedling protection and weed control. The field weeding performance test of three operation modes, namely, flat-pressure comb gear, compression-swinging helical gear and flat-pressure swing switch, were carried out. The results showed that the weeding rate of the flat pressing comb operation mode was 48.69%, and the seedling injury rate was 3.47%; The weeding rate was 86.01% and the seedling injury rate was 15.84% in the press swing bevel gear operation mode; The weeding rate was 81.02% and the seedling injury rate was 4.26% in the flat pressure swing switching operation mode. The optimal comprehensive weeding performance of the flat-pressure and swing-and-pressure switching operation mode meets the requirements of rice field seedling protection and weeding operation.

Path planning is one of the most important procedures in weeding machines at present. However, the small-scale intelligent machine can often be confined to crop damage and low efficiency caused by the imperfect path planning of the weeding components. In this study, adaptive path planning was proposed for the seedling-avoiding weeding operation using a heuristic search strategy. The soil-cultured lettuce was utilized as the experimental subject. A bidirectional parallel drive mechanism was concurrently designed to enhance the functionality of the weeding machine. A weeding platform was then constructed to integrate an image acquisition with the bidirectional parallel drive mechanism. The Denavit-Hartenberg (D-H) parameters were employed to perform the kinematic analysis. The precise control was realized for the movement of the weeding components. The crop damage was minimized for the effective removal of the weed. The YOLOv8n-seg model was applied to accurately identify and then extract the edge morphology of the lettuce and weeds from the acquired images. A minimum enclosing circle was delineated to define the lettuce protection zone. While the positions of the weeds were represented, according to their centroid coordinates. As such, clear and concise mapping was obtained for the weed locations in the field. According to the weed positions, adaptive recursive clustering was utilized to partition the weed locations into distinct clusters. A weighted undirected graph was then constructed for the operation path of the weeding component. The centers of the cluster also served as the key points in the graph. This graph-based approach was integrated with the A* (A-Star) heuristic search strategy. The shortest path planning was successfully achieved in the weeding components under seedling-avoiding conditions. The weeding machine was efficiently navigated around the protected lettuce areas to target weed clusters. The simulation was conducted to validate the effectiveness and adaptability of the path planning under various weed distributions. Additionally, a series of experiments were also carried out to compare with the conventional path planning under the identical weed distribution. The simulation results demonstrated that the optimal path planning was achieved in the coverage of the weed growth areas with the shorter paths across different weed distributions, indicating its strong adaptability and efficiency. The lengths of the operation path were then ranked in the ascending order of: the proposed, the Spiral, the Zigzag, and the BCD algorithm. Specifically, the path planning outperformed the rest under discrete weed distribution with grid coverage rates of 20%, 40%, and 60%. Furthermore, the path lengths were reduced by 54.6%, 53.3%, and 60.5%, respectively, compared with the BCD algorithm. The reduced length of the path enhanced the operational efficiency to minimize fuel consumption and wear and tear on the weeding machinery. Experimental results further verified the simulation findings. The adaptive seedling-avoiding weeding was achieved at a weeding rate of 93.91%, indicating a high-level removal of the weed. Importantly, the crop damage rate was maintained at 0, indicating the high precision to protect the cultivated lettuce from unintended harm. Additionally, the average weeding duration per unit (defined by the spatially adjacent lettuce plants in the row and column directions) was measured to be 14.2 s, indicating the high operational speed. The adaptive mode enhanced the operational efficiency by 42.28%, compared with the seedling-avoiding exhaustive weeding. There was promising potential for practical application in agricultural settings. The advanced image processing was integrated with precise kinematic control and path planning algorithms. A robust framework was then provided to improve the weeding performance of the small-scale intelligent machines. The heuristic search strategies were combined with adaptive clustering and precise motion control. The adaptive path planning of the seedling-avoiding weeding can be expected to reduce crop damage for high efficiency during mechanical weeding. In conclusion, the findings can offer a valuable technical reference for high-quality mechanical weeding in the field. The significant promise can greatly contribute to more sustainable and productive farming using intelligent weeding machinery.

A weed recognition was proposed using combinatorial deep learning, in order to reduce the influencing factors, such as the light shading of rice plants, interference of rice field algae and weeds with small targets. Data enhancement of weed sample was used to improve the model training and generalization for less overfitting. The MSRCP was introduced to enhance the image quality in complex environments. Weed target recognition was realized in the low contrast and clarity of rice field images, due to the light blockage of the rice plant. Front-end slicing and ViT (vision of transformer) classification were performed on the HD images. Information loss was avoided to detect the images in the process of network compression input. Small targets were retained in the high-definition images, in order to improve the effectiveness of the model in complex environments. The YOLOv7 model was replaced by the lightweight network GhostNet, and then embedded by the CA attention mechanism. The number of parameters and computations was reduced to enhance the feature extraction, particularly for the high accuracy and real-time performance of weed recognition. After the classification of target recognition, the image compression was attributed to the small target blurring and loss of effective information that was caused by only the image recognition model. The experiment showed that the weed dataset was expanded to improve the recognition of the model. The ablation test showed that the average mean accuracy of the test set after data enhancement was 84.9%, which was 10.8 percentage points better than the model trained on the original dataset. The ViT classification network outperformed Resnet 50 and Vgg, in terms of accuracy, recall and detection speed. Among them, the accuracy rate increased by 7.9 and 7.5 percentage points, respectively, and the recall rate increased by 7.1 and 5.3 percentage points, respectively. Comparative tests showed that the ViT network also achieved high classification accuracy and speed. The ablation test showed that the mean average accuracy of the improved YOLOv7 model was 88.2%, which was 3.3 percentage points higher than that of the original model. The number of parameters and the amount of computation were reduced by 10.43 M, and 66.54 ×10⁹ times/s, respectively, indicating the high speed and accuracy. The mean average accuracy of the improved model increased by 2.6 percentage points after MSRCP image enhancement before recognition. The mean average accuracy of the improved model increased by 5.3, 3.6, and 3.1 percentage points, respectively, for the light shading, algae interference, and similar shape of the rice leaf tip; Then, the mean average accuracy of the model further increased after the addition of ViT classification network. The mean accuracy also increased by 4.3 percentage points, compared with the original model. The mean average accuracies were improved by 6.2, 6.1, and 5.7 percentage points in three complex environments, respectively, compared with the original model. The mean average accuracy of the ViT-improved YOLOv7 model was 92.6%, which increased by 11.6, 10.1, 5.0, 4.2, and 4.4 percentage points, respectively, compared with the YOLOv5s, YOLOXs, MobilenetV3-YOLOv7, improved YOLOv7 and YOLOv8 model. Better detection was achieved to enable the weed identification in a complex rice field environment.