| As a crucial medium for transmitting information in nature,images play a vital role in information exchange and preservation,catering to the needs of various industries.However,challenges such as unbalanced image datasets,insufficient data,and poor quality in some fields reduce the usability of images,making image generation techniques a popular research topic.With the advancement of deep learning technology,image generation methods based on it have emerged.However,the existing methods encounter issues such as slow model convergence and poor quality due to the randomness of generated images,which pose difficulties for image generation.To address these challenges,this paper proposes a new image generation network model based on existing methods,comprising three modules: encoder,generator,and discriminator.The encoder encodes the training set images to obtain feature representation,which is then fed into the generator and discriminator networks for adversarial learning,ultimately reducing the randomness of the image generation algorithm,preventing network collapse,and enhancing image generation quality.The three primary research areas are as follows:(1)The T-AE-DCGAN network and the T-VAE-DCGAN network are proposed for the coding feature extraction module,respectively.Firstly,the auto-encoder(AE)is introduced as the coding module to form the T-AEDCGAN network to realize the image dimensionality reduction feature representation,guiding the learning and gaming process of the DCGAN network through the feature vector to reduce the network-generated images’ randomness and improve their usability.However,the output after auto-encoding lacks interpretable and exploitable structure.Therefore,this paper further proposes an adversarial network model,T-VAE-DCGAN,based on VAE encoding generation.The input data undergoes VAE encoding to obtain a hidden space vector that matches a specific distribution,which is then sampled as the generator’s input to enhance the model training’s convergence.The results on the public dataset demonstrate the method’s effectiveness.(2)Research for coding features and generator network fusion methods.Utilizing the generator’s multi-feature extraction layers,the fusion network structure model C-VAE-DCGAN is proposed.The feature representations encoded by VAE are randomly sampled and fused into each feature extraction layer of the generator,ensuring that all generator feature extraction layers contain the training set’s feature vectors.The comparative results indicate that the method can further accelerate model convergence and enhance generated image quality.(3)The fusion network is applied to an intelligent graphic drawing platform to create a demonstration system with intelligent image generation functionality.The platform accurately recognizes the output images from this paper’s model and controls the mechanical shafts to draw graphics,proving the practical application value of this paper’s method. |
