Research On Lossy Image Compression Method For Saliency Foreground Target Enhancement

With the advent of the information age,the use of images or video for information exchange has become the norm.As the resolution of images increases,it poses a huge challenge to the storage space of intelligent electronic devices,network bandwidth and other hardware facilities.Therefore,it is necessary to compress images in order to save storage space,reduce bandwidth usage and speed up image transmission.Image compression can reduce the image size to a reasonable range while ensuring the visual effect and quality of the image.Traditional image compression algorithms are difficult to achieve a balance between image compression ratio and image quality.According to the characteristics of the human eye visual system,the human eye will give priority to the saliency foreground target with limited attention,while the saliency detection method based on deep learning can use a large amount of training data to train a model with high detection accuracy,and improve the detection accuracy of salient foreground targets.The thesis proposes a lossy image compression algorithm for salient foreground target enhancement through an in-depth study of salient foreground target detection methods and image compression techniques,and uses a multi-scale salient target detection method based on adaptive perceptual fields to extract salient foreground targets from images.Based on this,a lossy image compression algorithm for foreground target enhancement is designed to achieve an increased compression ratio while improving the visual quality of the image.The main contents of the thesis are as follows:To address the problems of using fixed-size convolutional kernels to extract features in saliency foreground target detection methods,the perceptual field is restricted,feature extraction is inadequate,global semantic information extraction is inadequate,the detection effect after fusion of different features is poor,and the same processing method is used for features in different layers,the thesis divides the input features into three layers: low,medium and high,and uses different processing methods for features in different layers,and in the global semantic information extraction,a method based on dense cavity space pyramidal pooling is proposed to make the extracted information more continuous;in the feature map optimization,a residual refinement module is proposed to enable better fusion of features in different layers.The experimental results show that the F-measure curves of the proposed method have certain advantages over the other eight saliency detection algorithm on the ECSSD and HKU-IS datasets,and can suppress noise well in visual comparisons.To address the problem of unbalanced precision and recall of the saliency foreground target detection network,the initial saliency map obtained in the saliency foreground target detection method is optimized.The current optimization of the initial saliency map has the problem of a single optimization method and a complex optimization method,and a saliency map optimization method based on image fusion is proposed.The method first uses two optimization methods with different advantages to optimize the initial saliency map separately,and then fuses the two optimized saliency maps to obtain the final optimized saliency map.The experimental results show that the performance of the optimized model is improved,the generalization ability is enhanced,and the imbalance between precision and recall can be well resolved,especially in the PASCAL-S dataset.A lossy image compression method with saliency foreground target enhancement is proposed to address the imbalance between image compression ratio and image quality.This method adds a salient foreground target to the coding side of the image compression model.By calculating the importance score,the salient foreground target features are fused with the original image features to enhance the salient foreground targets,ensuring the quality of salient foreground target reconstruction,improve the overall quality of the reconstructed image.The experimental results show that the proposed method achieves good results in terms of visual quality when compared with JPEG,JPEG2000 and deep learning-based image compression algorithms,with clearer foreground images and richer details in the reconstructed images.