Research On Infrared Image Super-Resolution Algorithm Based On Adder Neural Networks

Infrared imaging is a technology that uses infrared rays to perceive the radiation and reflection of objects and generate images.Infrared radiation has various advantages,such as temperature sensitivity,day-night visibility and not affected by bad weather,which make it highly valuable in military weaponry,medical diagnosis,and civilian monitoring fields.However,due to the limitations of the materials and manufacturing processes of infrared detectors,the resulting infrared images are low in resolution,with unclear gray levels,blurry details,and poor visual effects.Benefited from the rapid development of deep learning technology,neural network-based super-resolution algorithms have significant advantages over traditional algorithms in objective indicators and visual effects in the field of high-resolution image generation.However,the huge computing resources required by neural network-based super-resolution networks,which make it difficult to be applied to mobile devices and hardware platforms.In order to improve the image quality of infrared images while reducing their computational resources and power consumption,research has been conducted around the infrared image super-resolution algorithm based on adder neural networks.The main research contents are as follows:(1)The thesis analyzes the characteristics of traditional super-resolution networks and adder neural networks.The VDSR network structure is used,which continuously repeats the same network layer combination,and the adder block is used to replace the convolutional layer in the network to reduce the multiplication operation in the network.To address the problems of performance loss and slow convergence caused by directly replacing the addition block,a new residual block is designed to place the batch normalization layer and activation function before the adder block,and the gaussian error unit activation function is introduced to obtain a lightweight network for infrared image super-resolution.Finally,ablation experiments were conducted on the network to verify the effectiveness and necessity of the above improvements.Under the condition of achieving a significant reduction in computational complexity and a 2.3times decrease in power consumption in the network,the performance loss is relatively small,and the visual effect is comparable to that of the VDSR network.(2)To address the problem of the large number of anomalous points in infrared images that cause significant oscillations in the model training process,the Huber Loss function was used as the loss function to reduce the impact of anomalous points on the training.In addition,the total variation loss function was added as a regularization term to remove noise while preserving the local details of the image,and to improve the overfitting problem caused by similar noise distributions in the same dataset during infrared image training,thereby increasing the model’s generalization ability.The results show that after improving the loss function,the infrared images obtained by the network have improved both objective indicators and visual effects on multiple datasets,reaching a level comparable to the VDSR network.(3)To address the issues of low contrast,poor grayscale level,and unsatisfactory visual effects in infrared images,a gradient adaptive sharpening technique and adaptive histogram equalization technique were proposed for preprocessing the image dataset.The dual-path infrared preprocessing process was then designed to separately preprocess low-resolution and high-resolution images to obtain residual images containing more high-frequency information.The model trained on the preprocessed dataset was able to enhance image contrast while preserving local detail information and without introducing a large amount of noise,resulting in a significant improvement in the visual effect of the reconstructed image.