To address the challenges of scale variation and dense object distribution in remote sensing imagery caused by varying imaging angles，a novel object detection algorithm is proposed based on multi-branch fusion self-attention（MFS）.A multi-branch module that integrates convolutional and self-attention mechanisms is designed to build a feature extraction network，and the fourth detection head is built for small objects to facilitate multi-scale feature fusion.Meanwhile，the resulting model is pruned by the DepGraph method to achieve a lightweight architecture.Experiments on the DOTA and NWPU VHR-10 datasets demonstrate that the proposed algorithm achieves mean average precision（mAP）scores of 77.7%and 96.5%respectively，outperforming the peer detectors of similar algorithm complexity.Notably，the pruned version maintains a mAP of 72.9%on DOTA，with only 6.64 million parameters.

To address the issues of missed detection and low detection accuracy in X-ray weld defect detection，an improved YOLOv8-based detection method is proposed.Firstly，the efficient multi-scale attention（EMA）mechanism is improved by replacing the 3×3 convolutional kernel with a 5×5 kernel to expand the receptive field，and replacing the average pooling with the multi-scale pooling to extract multi-scale features.The improved EMA module is embedded into the backbone network to enhance the model's ability to detect defects at various scales.Then the spatial pyramid pooling fast module is improved by introducing adaptive average pooling and max pooling layers，to improve the perception of weld edge information.Finally，in the neck part，Dual convolution is used to replace traditional convolution，to reduce the parameter number of the model.The WIoU（wise intersection over union）loss function is adopted to replace the CIoU（complete intersection over union）loss function to improve the convergence speed of the model. Experimental results show that，compared to YOLOv8n，the proposed algorithm reduces the number of parameters by 4.02%and increases the mean average precision by 5.9%，which is well-suited for X-ray weld defect detection tasks.

To address the challenges of small inter-class differences in pest details，severe field background interference，and imbalanced sample distribution，a complementary feature fusion dual-stream network for pest recognition is proposed.This network combines the local perception capability of convolutional neural networks with the global modeling ability of the Mamba model，capturing and integrating the global and local information of pest images.A hierarchical multiscale perception module is designed to extract multi-scale image features through grouped hierarchical convolution and enhance pest detail information with a detail enhancement perception strategy.An adaptive focusing Mamba module is designed to locate key pest regions using dynamic convolution operators and reduce complex background interference.Additionally，an attention-weighted fusion module is designed to achieve adaptive interaction and optimization of global and local features through a cross-attention mechanism，further improving the accuracy of semantic expression.A balanced loss function is constructed to mitigate the effects of class imbalance in the dataset.The experimental results show that the network achieves an accuracy of 71.19%on the large-scale pest dataset IP102，and an accuracy of 99.36%on the D0 dataset，demonstrating its ability to effectively identify pest species.

For the low detection accuracy problem of small targets and low-contrast defects on the surface of latex gloves，an improved YOLOv（you only look once version）8n algorithm for defects detection on the surface of latex gloves is proposed.The receptive field attention convolution module is introduced in the feature extraction network to dynamically adjust the spatial feature weights within the receptive field，and to enhance the network's focus on defect features. The C2f module is redesigned based on the proposed multi-scale convolution，which captures the contextual information from shallow features through multi-scale convolutional kernels，and improves the network's ability to extract shallow features.The context and the spatial feature calibration network are added to the feature fusion network，where feature calibration refines and aligns contextual information and spatial features，and further enhances the representation of defect features.Experimental results show that on the homemade dataset，the mean average precision（mAP）of the improved algorithm reaches 93.2%，which is 3.1%higher than that of YOLOv8n.It effectively improves the surface defect detection accuracy of latex gloves.In addition，on the VisDrone2019Det and steel defect detection datasets，the mAP reaches 36.1%and 79.8%，respectively，which are 1.1%and 2.7%higher than that of YOLOv8n，and further verifies the effectiveness of the improved algorithm.