Adversarial patch is an important form of real-world adversarial attack that brings serious risks to the robustness of deep neural networks. Previous methods generate adversarial patches by either optimizing their perturbation values while fixing the pasting position or manipulating the position while fixing the patch's content. This reveals that the positions and perturbations are both important to the adversarial attack. For that, in this paper, we propose a novel method to simultaneously optimize the position and perturbation for an adversarial patch, and thus obtain a high attack success rate in the black-box setting. Technically, we regard the patch's position, the pre-designed hyper-parameters to determine the patch's perturbations as the variables, and utilize the reinforcement learning framework to simultaneously solve for the optimal solution based on the rewards obtained from the target model with a small number of queries. Extensive experiments are conducted on the Face Recognition (FR) task, and results on four representative FR models show that our method can significantly improve the attack success rate and query efficiency. Besides, experiments on the commercial FR service and physical environments confirm its practical application value. We also extend our method to the traffic sign recognition task to verify its generalization ability.

The accurate and rapid acquisition of crop and weed information is an important prerequisite for automated weeding operations. This paper proposes the application of a network model based on Faster R-CNN for weed identification in images of cropping areas. The feature pyramid network (FPN) algorithm is integrated into the Faster R-CNN network to improve recognition accuracy. The Faster R-CNN deep learning network model is used to share convolution features, and the ResNeXt network is fused with FPN for feature extractions. Tests using >3000 images for training and >1000 images for testing demonstrate a recognition accuracy of >95%. The proposed method can effectively detect weeds in images with complex backgrounds taken in the field, thereby facilitating accurate automated weed control systems.

One of the most important food crops is rice. For this reason, the accurate identification of rice pests is a critical foundation for rice pest control. In this study, we propose an algorithm for automatic rice pest identification and classification based on fully convolutional networks (FCNs) and select 10 rice pests for experiments. First, we introduce a new encoder–decoder in the FCN and a series of sub-networks connected by jump paths that combine long jumps and shortcut connections for accurate and fine-grained insect boundary detection. Secondly, the network also integrates a conditional random field (CRF) module for insect contour refinement and boundary localization, and finally, a novel DenseNet framework that introduces an attention mechanism (ECA) is proposed to focus on extracting insect edge features for effective rice pest classification. The proposed model was tested on the data set collected in this paper, and the final recognition accuracy was 98.28%. Compared with the other four models in the paper, the proposed model in this paper is more accurate, faster, and has good robustness; meanwhile, it can be demonstrated from our results that effective segmentation of insect images before classification can improve the detection performance of deep-learning-based classification systems.

In modern animal husbandry, intelligent digital farming has become the key to improve production efficiency. This paper introduces a model based on improved YOLOv8, Cattle Behavior Recognition-YOLO (CBR-YOLO), which aims to accurately identify the behavior of cattle. We not only generate a variety of weather conditions, but also introduce multi-target detection technology to achieve comprehensive monitoring of cattle and their status. We introduce Inner-MPDIoU Loss and we have innovatively designed the Multi-Convolutional Focused Pyramid module to explore and learn in depth the detailed features of cattle in different states. Meanwhile, the Lightweight Multi-Scale Feature Fusion Detection Head module is proposed to take advantage of deep convolution, achieving a lightweight network architecture and effectively reducing redundant information. Experimental results prove that our method achieves an average accuracy of 90.2% with a reduction of 3.9 G floating-point numbers, an increase of 7.4%, significantly better than 12 kinds of SOTA object detection models. By deploying our approach on monitoring computers on farms, we expect to advance the development of automated cattle monitoring systems to improve animal welfare and farm management.

Precise weed recognition is an important step towards achieving intelligent agriculture. In this paper, a novel weed recognition model, Cotton Weed-YOLO, is proposed to improve the accuracy and efficiency of weed detection. CW-YOLO is based on YOLOv8 and introduces a dual-branch structure combining a Vision Transformer and a Convolutional Neural Network to address the problems of the small receptive field of the CNN and the high computational complexity of the transformer. The Receptive Field Enhancement (RFE) module is proposed to enable the feature pyramid network to adapt to the feature information of different receptive fields. A Scale-Invariant Shared Convolutional Detection (SSCD) head is proposed to fully utilize the advantages of shared convolution and significantly reduce the number of parameters in the detection head. The experimental results show that the CW-YOLO model outperforms existing methods in terms of detection accuracy and speed. Compared with the original YOLOv8n, the detection accuracy, mAP value, and recall rate are improved by 1.45, 0.7, and 0.6%, respectively, the floating-point numbers are reduced by 2.5 G, and the number of parameters is reduced by 1.52 × 10⁶ times. The proposed CW-YOLO model provides powerful technical support for smart agriculture and is expected to promote the development of agricultural production in the direction of intelligence and precision.

This paper uses an enhanced sparrow search algorithm (SSA) to optimise the support vector machine (SVM) by considering the emission of air pollution sources as the independent variable. Consequently, it establishes a PM2.5 concentration prediction model to improve the prediction accuracy of fine particulate matter PM2.5 concentration. First, the principal component analysis is applied to extract key variables affecting air quality from high-dimensional air data to train the model while removing unnecessary redundant variables. Adaptive dynamic weight factors are introduced to balance the global and local search capabilities and accelerate the convergence of the SSA. Second, the SSA–SVM prediction model is defined using the optimised SSA to continuously update the network parameters and achieve the rapid prediction of atmospheric PM2.5 concentration. The findings demonstrate that the optimised SSA–SVM prediction method can quickly predict atmospheric PM2.5 concentration, using the cyclic search method for the best solution to update the model, proving the method’s effectiveness. Compared with other methods, this approach has a small prediction error, a high prediction accuracy and better practical value.