The key to precise weeding in the field lies in the efficient detection of weeds. There are no studies on weed detection in potato fields. In view of the difficulties brought by the cross-growth of potatoes and weeds to the detection of weeds, the existing detection methods cannot meet the requirements of detection speed and detection accuracy at the same time. This study proposes an improved YOLOv4 model for weed detection in potato fields. The proposed algorithm replaces the backbone network CSPDarknet53 in the YOLOv4 network structure with the lightweight MobileNetV3 network and introduces Depthwise separable convolutions instead of partial traditional convolutions in the Path Aggregation Network (PANet), which reduces the computational cost of the model and speeds up its detection. In order to improve the detection accuracy, the convolutional block attention module (CBAM) is fused into the PANet structure, and the CBAM will process the input feature map with a channel attention mechanism (CAM) and spatial attention mechanism (SAM), respectively, which can enhance the extraction of useful feature information. The K-means++ clustering algorithm is used instead of the K-means clustering algorithm to update the anchor box information of the model so that the anchor boxes are more suitable for the datasets in this study. Various image processing methods such as CLAHE, MSR, SSR, and gamma are used to increase the robustness of the model, which eliminates the problem of overfitting. CIoU is used as the loss function, and the cosine annealing decay method is used to adjust the learning rate to make the model converge faster. Based on the above-improved methods, we propose the MC-YOLOv4 model. The mAP value of the MC-YOLOv4 model in weed detection in the potato field was 98.52%, which was 3.2%, 4.48%, 2.32%, 0.06%, and 19.86% higher than YOLOv4, YOLOv4-tiny, Faster R-CNN, YOLOv5 l, and SSD(MobilenetV2), respectively, and the average detection time of a single image was 12.49ms. The results show that the optimized method proposed in this paper outperforms other commonly used target detection models in terms of model footprint, detection time consumption, and detection accuracy. This paper can provide a feasible real-time weed identification method for the system of precise weeding in potato fields with limited hardware resources. This model also provides a reference for the efficient detection of weeds in other crop fields and provides theoretical and technical support for the automatic control of weeds.

To guide grape picking robots to recognize and classify the grapes with different maturity quickly and accurately in the complex environment of the orchard, and to obtain the spatial position information of the grape clusters, an algorithm of grape maturity detection and visual pre-positioning based on improved YOLOv4 is proposed in this study. The detection algorithm uses Mobilenetv3 as the backbone feature extraction network, uses deep separable convolution instead of ordinary convolution, and uses the h-swish function instead of the swish function to reduce the number of model parameters and improve the detection speed of the model. At the same time, the SENet attention mechanism is added to the model to improve the detection accuracy, and finally the SM-YOLOv4 algorithm based on improved YOLOv4 is constructed. The experimental results of maturity detection showed that the overall average accuracy of the trained SM-YOLOv4 target detection algorithm under the verification set reached 93.52%, and the average detection time was 10.82 ms. Obtaining the spatial position of grape clusters is a grape cluster pre-positioning method based on binocular stereo vision. In the pre-positioning experiment, the maximum error was 32 mm, the mean error was 27 mm, and the mean error ratio was 3.89%. Compared with YOLOv5, YOLOv4-Tiny, Faster_R-CNN, and other target detection algorithms, which have greater advantages in accuracy and speed, have good robustness and real-time performance in the actual orchard complex environment, and can simultaneously meet the requirements of grape fruit maturity recognition accuracy and detection speed, as well as the visual pre-positioning requirements of grape picking robots in the orchard complex environment. It can reliably indicate the growth stage of grapes, so as to complete the picking of grapes at the best time, and it can guide the robot to move to the picking position, which is a prerequisite for the precise picking of grapes in the complex environment of the orchard.