Search

In the task of knowledge distillation, a dual-source adaptive knowledge distillation (DSAKD) method is proposed to address the issues of feature information loss during the feature alignment process and the lack of consideration for the differences in samples in the soft label distillation method. The DSAKD method extracts more discriminative knowledge from both the feature layer and soft labels of the teacher network, which enhances the performance of the lightweight student network. An attention-based feature adaptive fusion module was proposed to integrate the intermediate layer features of the teacher network and the student network, and then the feature embedding contrastive distillation strategy was used to optimize the features of the student network. An adaptive temperature distillation strategy was also proposed, which assigned different temperature coefficients to all training samples adaptively based on the prediction confidence of the teacher network. The experimental results show that our proposed method achieves the optimal distillation effect on three benchmark datasets, significantly improving the classification performance of lightweight student networks. Specifically, compared with the best-performing method, the proposed method improves the average top-1 validation accuracy on CIFAR10, CIFAR100, and ImageNet datasets by 0.46%, 0.41%, and 0.59%, respectively.

The multi-stage deep neural network in the blind motion image deblurring task lacks a large range of receptive fields and it was difficult to reasonably interact with the image features of each stage. A progressive depth network (PDNet) with dilated convolution and contextual attention fusion module was thus proposed for restoring sharp images. This method included three stages:local feature extraction, image feature integration and image restoration. In the local feature extraction stage and image feature integration stage, multi-branch dilated convolution module (MDCB) was used to increase receptive field to adapt to different degrees of motion blur. In the image feature integration stage and image restoration stage, the contextual attention fusion module was used for information interaction of image features in different stages to achieve progressive image feature enhancement. The proposed method could make full use of local and global image features to guide image restoration through the three-stage progressive enhancement strategy, so as to generate sharp and high-quality images. The experimental results showed that compared with SRN and other networks, the proposed PDNet had a better effect on GoPro data set and RealBlur-J data set, and the peak signal to noise ratio (PSNR) had an average increase of 2.9 dB, and the structural similarity (SSIM) had an average increase of 0.05.

In order to solve the problem of poor consistency between the traditional super-resolution image quality assessment (SRIQA) index and human subjective perception, this paper presents a reference free super-resolution image quality evaluation method by using multi-layer perceptron (MLP). This method used the pre trained VGG16 network to extract the perceptual quality features of the super-resolution image, and established the regression model between the perceptual quality features of the super-resolution image and the corresponding subjective quality score through MLP. Experimental results show that the Pearson linear correlation coefficient (PLCC) and Spearman rank order correlation coefficient (SROCC) of the proposed algorithm on the public super-resolution image data set exceed 0.95, which is significantly superior to other existing image quality evaluation methods, and has higher consistency with human subjective perception.