Virtual simulation of maintainability is an important tool for aircraft structure and systems design， in which the rapid generation of appropriate virtual human maintenance posture is crucial to the efficiency and feasibility of maintainability analysis. Aiming at the problems of low efficiency and limited applicability of the current virtual human maintenance posture solving methods， a virtual maintenance posture layered solving method for the narrow space of aircraft was proposed. In the method， with the waist as the demarcation point， the maintenance posture was decomposed into two parts： upper body and lower body， and Elitist Nondominated Sorting Genetic Algorithm Ⅱ （NSGA-Ⅱ） was used to optimize and solve them respectively to obtain the overall maintenance posture. Firstly， the maintenance posture optimization criterion was constructed by considering the space limitations and human body structure constraints comprehensively. Secondly， multi-objective optimization models of upper body and lower body postures were constructed on the basis of the maintenance posture optimization criterion through analyzing geometry and inverse kinematics principles， and NSGA-Ⅱ was applied to solve them sequentially. Through analysis of cases of disassembly of aircraft cockpit Air Data Module （ADM） and cargo promotion in cargo hold， it is verified that the proposed method can generate virtual maintenance posture in narrow space of aircraft effectively， and has good applicability and feasibility.

In response to the issue of low accuracy in handling complex work conditions in rolling bearing fault diagnosis， a Multi-Task Learning （MTL） model naming as Multi-pathed Hierarchical Mixture-of-Experts （MHMoE）， and the corresponding hierarchical training mode were proposed. In this model， by combining multi-stage， multi-task joint training， a hierarchical information sharing mode was achieved. The model's generalization and fault recognition accuracy were further improved on the basis of the ordinary MTL mode， enabling the model to perform tasks on both complex and simple datasets excellently. Meanwhile， by incorporating the bottleneck layer structure of one-dimensional ResNet， the depth of the network was ensured while avoiding issues such as vanishing and exploding gradients， so as to extract relevant features of the dataset fully. Experimental results on the Paderborn University bearing fault dataset （PU） as the test dataset demonstrate that under varying degrees of working complexity， compared to the OMoE （One-gate Mixture-of-Experts） -ResNet18 model without MTL， the proposed model has the accuracy improved by 5.45 to 9.30 percentage points. Compared to the models such as Ensemble Empirical Mode Decomposition Hilbert spectral transform （EEMD-Hilbert）， MMoE （Multi-gate Mixture-of-Experts）， and Multi-Scale multi-Task Attention Convolutional Neural Network （MSTACNN）， the proposed model has the accuracy improved by 3.21 to 16.45 percentage points at least.

In the Radio Frequency IDentification （RFID） real-time defect detection task in industrial scenes， the deep learning target detection algorithms such as You Only Look Once （YOLO） are often adopted in order to ensure the detection precision and speed. However， these algorithms are still difficult to meet the speed requirement of industrial detection， and the corresponding network models cannot be deployed on resource-constrained devices. To solve these problems， the YOLO model must be pruned and compressed. A new network pruning method of the weighted fusion of feature information richness and feature information diversity based on rank information was proposed. Firstly， the unpruned model was loaded and reasoned， and the rank information of the corresponding feature maps of the filters was obtained in forward propagation to measure the feature information richness. Secondly， according to the different pruning rates， the rank information was clustered or the similarity of the rank information was calculated to measure the feature information diversity. Finally， the importance degrees of the corresponding filters were obtained after the weighted fusion and were sorted， and the filters with low importance were cut off. Experimental results show that， for YOLOv4， when the pruning rate is 28.87% and the weight of feature information richness is 0.75， the proposed method has the mean Average Precision （mAP） improved by 2.6% - 8.9% compared with the method that uses rank information of the feature maps alone， and the model pruned by the proposed method even has the mAP increased by 0.4% and the model parameters reduced by 35.0% compared with the unpruned model， indicating that the proposed method is conducive to the model deployment.

Concerning the problem that the surface defects of the coin are small， variable in shape， easily confused with the background and difficult to be detected， an improved algorithm of coin surface defect detection named DCA-YOLO （Deformable Convolution and Adaptive space feature fusion-YOLO） was proposed. First of all， due to the different shapes of defects， three network structures with deformable convolution modules added at different positions in the backbone network were designed， and the ability to extract defects was improved through convolution learning offset and adjusting parameters. Then， the adaptive spatial feature fusion network was used to learn the weight parameters to better adapt to targets with different scales by adjusting the contribution of each pixel in the feature maps of different scales. Finally， the anchor ratio was adjusted， the category weights were dynamically adjusted， the comparison network performance was optimized， thus， a model network to add deformable convolution before upsampling for multi-scale fusion of the output features of the backbone network was proposed. Experimental results show that on the coin defect dataset， the detection mAP （mean Average Precision） of DCA-YOLO algorithm reaches 92.8%， which is close to that of Faster-RCNN （Faster Region-based Convolutional Neural Network）； compared with YOLOv3， the proposed algorithm has the detection speed basically the same with 3.3 percentage points improvement on detection mAP， and 3.2 percentage points increase on F1-score.

In production and life， the existence of motion blur increases the difficulty of Quick Response code （QR code） recognition. To solve this problem， a motion blur removal algorithm for QR code images based on blur kernel estimation and alternating Transformer was proposed. Firstly， in order to solve the problem that the current motion blur removal algorithms lack explanation of the intermediate degradation process， a blur Kernel Estimation Network （KEN） was used to estimate the shapes and parameters of the blur kernel dynamically， and after performing Wiener filtering on KEN output and the original image， the subsequent restoration networks were guided to better remove motion blur. Then， aiming at the problems that the window-based Transformer has a weak ability to capture global features and the traditional Transformer has high computational complexity， an Alternating Transformer Block （ATB） that combines Local-window Transformer Block （LTB） and Global-axis Transformer Block （GTB） was proposed to extract local and global features alternately. Finally， since when the input is a single-scale image， the model cannot handle with different levels of blur， a Multi-Scale Feature Fusion Block （MSFFB） was proposed. In this way， the model was able to extract features from multi-scale input images， utilize contextual information effectively， and retain and restore image details better. Experimental results on a motion blurred QR code image dataset show that for the linear blur kernel test set， compared with Uformer （U-shaped Transformer）-B， which has the second best restoration effect， the proposed algorithm has better performance in Peak Signal-to-Noise Ratio （PSNR） and Structural SIMilarity （SSIM） with 3.11% and 1.23% improvements respectively； for the nonlinear blur kernel test set， compared with Uformer-B， the proposed algorithm has the PSNR and SSIM indicators increased by 7.13% and 2.19% respectively. At the same time， the Multiply ACcumulate operations （MAC） of the proposed algorithm is decreased by 77.22%， obtaining the best among all comparison algorithms， and the proposed algorithm has a decrease of 83.5% in the model Parameter （Param）. Besides， YOLOv4 and ZBar were used for object detection and recognition experiments， and the results show that the proposed algorithm has certain practical significance for improving the efficiency of QR code detection and recognition.

Optical anti-counterfeiting elements of diffractive optically variable image have been widely adopted in various high-security or high-value-added printed products such as banknotes， certificates/cards， and product packaging to prevent counterfeiting. This kind of optical anti-counterfeiting elements have optically variable characteristics， and three-dimensional and dynamically changing images generated by these elements under white light illumination conditions will cause serious interference to anomaly detection， further making it difficult to detect abnormal situations in industrial production process timely. In response to the challenge that the optical variable characteristics of diffractive optically variable images cannot be adapted by the existing unsupervised anomaly detection algorithms， a diffractive optically variable image anomaly detection algorithm based on Reverse Distillation （RD） was proposed. In the algorithm， noise was added to normal samples to generate pseudo-anomalous samples， and possible abnormal phenomena in industrial fields were simulated. Subsequently， both normal and pseudo-anomalous samples were input into the network as image pairs， and based on the Siamese network architecture， a Contrast and Reconstruction Module （CRM） was proposed. In this module， contrastive learning and reconstruction were conducted to the features extracted by the encoder from normal samples and pseudo-anomalous samples through feature reconstruction layer. This not only avoided the inflow of abnormal information into the decoder， leading to distillation failure， but also ensured that the reconstructed features conformed to the normal distribution of samples. Following this， the reconstructed features were input into feature fusion layer and feature compression layer for feature fusion and dimension reduction， and the compressed features were decoded layer by layer using the decoder. Finally， by using collaborative discrepancy optimization algorithm， the decoded features and the features extracted by the encoder were distilled to identify and locate abnormal information within the samples. Experimental results show that compared to the existing advanced anomaly detection algorithms on a certain anti-counterfeiting label dataset， the proposed algorithm improves adaptability to optically variable characteristics of diffractive optically variable images， and maintains high detection accuracy for abnormal regions within samples， achieving 100% in Image-Area Under Receiver Operator Curve （Image-AUROC）， 95.02% in Pixel-Area Under Receiver Operator Curve （Pixel-AUROC）， and 92.98% in Per-Region-Overlap （PRO）. These results meet the requirements for anomaly detection of diffractive optically variable images in industrial fields.

Among the classic tasks in machine vision， semantic segmentation is a category with a large amount of calculation， making it difficult to deploy Convolutional Neural Networks （CNNs） for segmentation in edge computing systems. Field Programmable Gate Array （FPGA） is a hardware widely used in industrial vision sensors for data stream processing. In recent years， methods for deploying CNNs on FPGA have been proposed. However， due to limited computing resources， current technology cannot achieve acceptable speed and accuracy when performing semantic segmentation of high-resolution images on FPGA. After analyzing the characteristics of deep learning accelerators on FPGA， a new segmentation network， Trilateral Segment Network （TriSeNet）， was proposed to achieve end-to-end inference of semantic segmentation tasks of high-resolution images on edge accelerators. TriSeNet was deployed on Xilinx Kria K26 SOM to process CityScapes semantic segmentation. TriSeNet achieved a mean Intersection over Union （mIoU） of 75%； for images with resolution of 512*1 024，it had a inference speed of 32 FPS. It could utilize computing resources at the edge efficiently， and achieved a calculator utilization of 62.6%. It is verified that TriSeNet is a model adapting to hardware characteristics of the accelerator successfully.