| Deep learning neural network models are developing rapidly in the field of computer vision as theories and techniques continue to improve.Transformer,a network model first used in natural language processing,has achieved great success in semantic analysis due to its powerful modelling,global information interaction and parallel computing capabilities.Researchers next applied Transformer to the field of computer vision.Through the encoder-decoder architecture and the global information extraction capabilities of the self-attentive mechanism,Transformer has also achieved good results in computer vision tasks such as image classification and target detection,demonstrating the model’s great potential.At the same time,a number of issues have arisen as researchers have migrated Transformer to the computer vision domain.Firstly,the change from sequential to 2D image inputs presents a drawback in terms of positional encoding,and secondly,the model is very slow to converge iteratively during training.As the detection efficiency and training of this type of model need more optimisation,thesis proposes different solutions to these problems.In order to solve the problem of position encoding caused by different dimensions of input format,thesis adds a more effective computational embedding of relative position encoding on the basis of absolute position encoding in ViT model by analyzing the similarities and differences between one-dimensional input and two-dimensional input modes.After experiments with various combinations of approaches,the directed relative position coding with the largest lifting space is added to the self-attentive computational process in the ViT model for interaction,allowing the model to model the features of the input images more accurately.Applying the method to the Dei T model and DETR model for image classification and object detection tasks resulted in a nearly 1% improvement in the detection accuracy of the model.To address the problem of slow model iteration,thesis recombines the decoder part of the computational formulae based on the task decoupling optimization of the multi-head attention matching computational link,resulting in a tenfold improvement in model iteration efficiency.The new object detection model NPET obtained after the combination of the above optimisation methods further modified the model,and obtained an accuracy of 45% on the COCO2017 datasets,which is nearly 3% higher than the original model DETR,proving the effectiveness of this improved method.On this basis,a visual target detection system based on the PyQt framework is designed in thesis.The visualisation results of different object detection models are realised,and a comparison between the results of two-stage and single-stage mainstream object detection models is shown.Real-time image classification and regression box prediction functions are implemented,and image upload,image storage and object detection functions are completed.The system was finally tested for functionality and performance,and its usability was verified. |
