Frequent outbreaks of agricultural pests can reduce crop production severely and restrict agricultural production. Therefore, automatic monitoring and precise recognition of crop pests have a high practical value in the process of agricultural planting. In recent years, pest recognition and detection have been rapidly improved with the development of deep learning-based methods. Although certain progress has been made in the research on pest detection and identification technology based on deep learning, there are still many problems in the production application in a field environment. This work presents a pest detector for multi-category dense and tiny pests named the Pest-YOLO. First, the idea of focal loss is introduced into the loss function using weight distribution to improve the attention of hard samples. In this way, the problems of hard samples arose from the uneven distribution of pest populations in a dataset and low discrimination features of small pests are relieved. Next, a non-Intersection over Union bounding box selection and suppression algorithm, the confluence strategy, is used. The confluence strategy can eliminate the errors and omissions of pest detection caused by occlusion, adhesion and unlabeling among tiny dense pest individuals to the greatest extent. The proposed Pest-YOLO model is verified on a large-scale pest image dataset, the Pest24, which includes more than 20k images with over 190k pests labeled by agricultural experts and categorized into 24 classes. Experimental results show that the Pest-YOLO can obtain 69.59% for mAP and 77.71% for mRecall on the 24-class pest dataset, which is 5.32% and 28.12% higher than the benchmark model YOLOv4. Meanwhile, our proposed model is superior to other several state-of-the-art methods, including the SSD, RetinaNet, Faster RCNN, YOLOv3, YOLOv4, YOLOv5s, YOLOv5m, YOLOX, DETR, TOOD, YOLOv3-W, and AF-RCNN detectors. The code of the proposed algorithm is available at: https://github.com/chr-secrect/Pest-YOLO.

The number of wheat spikes per unit area is one of the most important agronomic traits associated with wheat yield. However, quick and accurate detection for the counting of wheat spikes faces persistent challenges due to the complexity of wheat field conditions. This work has trained a RetinaNet (SpikeRetinaNet) based on several optimizations to detect and count wheat spikes efficiently. This RetinaNet consists of several improvements. First, a weighted bidirectional feature pyramid network (BiFPN) was introduced into the feature pyramid network (FPN) of RetinaNet, which could fuse multiscale features to recognize wheat spikes in different varieties and complicated environments. Then, to detect objects more efficiently, focal loss and attention modules were added. Finally, soft non-maximum suppression (Soft-NMS) was used to solve the occlusion problem. Based on these improvements, the new network detector was created and tested on the Global Wheat Head Detection (GWHD) dataset supplemented with wheat-wheatgrass spike detection (WSD) images. The WSD images were supplemented with new varieties of wheat, which makes the mixed dataset richer in species. The method of this study achieved 0.9262 for mAP50, which improved by 5.59, 49.06, 2.79, 1.35, and 7.26% compared to the state-of-the-art RetinaNet, single-shot multiBox detector (SSD), You Only Look Once version3 (Yolov3), You Only Look Once version4 (Yolov4), and faster region-based convolutional neural network (Faster-RCNN), respectively. In addition, the counting accuracy reached 0.9288, which was improved from other methods as well. Our implementation code and partial validation data are available at https://github.com/wujians122/The-Wheat-Spikes-Detecting-and-Counting.

Early blight and late blight are important factors restricting tomato yield. However, it is still a challenge to accurately and objectively detect and segment crop diseases in order to evaluate disease damage. In this paper, the Disease Segmentation Detection Transformer (DS-DETR) is proposed to segment leaf disease spots efficiently based on several improvements to DETR. Additionally, a damage assessment is carried out by the area ratio of the segmented leaves to the disease spots. First, an unsupervised pre-training method was introduced into DETR with the Plant Disease Classification Dataset (PDCD) to solve the problem of the long training epochs and slow convergence speed of DETR. This method can train the Transformer structures in advance to obtain leaf disease features. Loading the pre-training model weight in DS-DETR can speed up the convergence speed of the model. Then, Spatially Modulated Co-Attention (SMCA) was used to assign Gaussian-like spatial weights to the query box of DS-DETR. The different positions in the image are trained using the query boxes with different weights to improve the accuracy of the model. Finally, an improved relative position code was added to the Transformer structure of DS-DETR. Relative position coding promotes the capture of the sequence order of input tokens by the Transformer. The spatial location feature is strengthened by establishing the location relationship between different instances. Based on these improvements, the DS-DETR model was tested on the Tomato leaf Disease Segmentation Dataset (TDSD) constructed by us. The experimental results show that the DS-DETR proposed by us achieved 0.6823 for AP_mask, which improved by 12.87%, 8.25%, 3.67%, 1.95%, 10.27%, and 9.52% compared with the state-of-the-art: Mask RCNN, BlendMask, CondInst, SOLOv2, ISTR, and DETR, respectively. In addition, the disease grading accuracy reached 0.9640 according to the segmentation results given by our proposed model.

In complex citrus orchard environments, light changes, branch shading, and fruit overlapping impact citrus detection accuracy. This paper proposes the citrus detection model YOLO-DCA in complex environments based on the YOLOv7-tiny model. We used depth-separable convolution (DWConv) to replace the ordinary convolution in ELAN, which reduces the number of parameters of the model; we embedded coordinate attention (CA) into the convolution to make it a coordinate attention convolution (CAConv) to replace the ordinary convolution of the neck network convolution; and we used a dynamic detection head to replace the original detection head. We trained and evaluated the test model using a homemade citrus dataset. The model size is 4.5 MB, the number of parameters is 2.1 M, mAP is 96.98%, and the detection time of a single image is 5.9 ms, which is higher than in similar models. In the application test, it has a better detection effect on citrus in occlusion, light transformation, and motion change scenes. The model has the advantages of high detection accuracy, small model space occupation, easy application deployment, and strong robustness, which can help citrus-picking robots and improve their intelligence level.

Semantic segmentation of large-scale point clouds in 3D computer vision is a challenging problem. Existing feature extraction modules often emphasize learning local geometry while not giving adequate consideration to the integration of color information. This limitation prevents the network from thoroughly learning local features, thereby impacting segmentation accuracy. In this study, we propose three modules for robust feature extraction and aggregation, forming a novel point cloud segmentation network (HFA-Net) for large-scale point cloud semantic segmentation. First, we introduce the Hybrid Feature Extraction Component (HFEC) and the Hybrid Bilateral Enhancement Component (HBAC) to comprehensively extract and enhance the geometric, color, and semantic information of point clouds. Second, we incorporate the Ternary-Distance Attention Pooling (TDAP) module, which leverages trilateral distances to further refine the network’s focus on various features, enabling it to emphasize both locally important features and broader local neighborhoods. These modules are stacked into dense residual components to expand the network’s receptive field. Our experiments on several large-scale benchmark datasets, including Semantic3D, Toronto3D, S3DIS and LASDU demonstrate the effectiveness of HFA-Net when compared to state-of-the-art networks.