Research Of Image Enhancement Methods For Multi-stage Feature Fusion

Along with the development of the computer science field of computer vision,people more and more high to the requirement of image quality,whether it’s a wild night or weak indoor light environment,the underexposure collected images or video in the scene,often because of the brightness and contrast of different difficult to distinguish between the detail of the image,and image of the low-light environment often have a lot of noise,the obtained data can’t meet people’s expectations,nor can it be applied to the needs of advanced visual tasks.Low-light image enhancement can meet people’s needs on image brightness,contrast and color saturation after digital image processing.Its core idea is to effectively improve brightness and contrast while restoring its details and color,and avoid some noise caused by low illumination environment.Low illumination image enhancement technology is widely used in field military reconnaissance,anti-theft monitoring at night,image backlight processing and automatic driving at night.This paper mainly discusses the performance of deep neural network in single image processing task for low illumination image enhancement task.By designing image enhancement network,the experiment is further extended to image rain removal task and underwater image enhancement task,and some classical theories and algorithms are further discussed.The main contributions of this paper are as follows:1.Based on the theory that the brain retina Retinex theory approach,according to the traditional Retinex theory about the decomposition of the image of complicated operation,combined with multi-branch neural networks improved and designed a kind of end-to-end neural network structure,through the use of the existing neural network model as a branch network,to extract the image reflection components;at the same time,the attention module is used to process the different branch results,and the obtained features are spliced according to the channel dimension,and finally the reflection component is obtained.In the image restoration module,U-net is used to fuse the reflection and illumination components of the image,and finally get the enhanced image.All these operations are completed in the same neural network model,avoiding setting intermediate parameters by manual experience.The results show that this algorithm can extract the reflection information of the image effectively,and has excellent effect on the brightness,contrast and color of the image.2.In the image enhancement algorithm based on model stacking,the main purpose is to design a method that can adapt to the real-time requirements of image processing,so as to better apply in automatic driving and other technologies.In this paper,we design a new model of stack structure,no longer by using traditional residual network or u-shaped network model,using multiple convolution layer to the input feature extraction,and then into this section of the proposed model is stacked in the network,after the fusion of each module are characteristics,finally the final image is obtained by image recovery module.Through a large number of experiments,the method presented in this master’s thesis can adapt to the requirements of real-time performance and achieve excellent performance in both subjective and objective indicators.3.In the method based on multi-level feature fusion,the main purpose is to be able to process images under various task scenarios,including rain line images and underwater images.The method of multi-stage feature fusion is to process the image in stages.After the image is processed by multi-stage module,it is sent to the next level module for further processing,and the final result is obtained by gradually optimizing the image structure.In the multi-stage feature fusion module,the first is the image feature pre-extraction part composed of a gated neural network and two convolutional neural networks,followed by a three-branch image processing part,in which this master’s thesis designed a three-channel attention mechanism,through this method to further image processing.At the same time,by referring to the result features of the upper level in the process of processing,this can better reference the image context features.The results show that the proposed method can perform well in low illumination image enhancement task,image rain removal task and underwater image enhancement task,and has good performance in both subjective and objective aspects,and can meet the needs of various image restoration figures.