| The widespread application of neural networks has brought us a lot of convenience,but some problems that come with it urgently need to be solved.Firstly,deeper models require a large number of samples as input,and when the data volume cannot meet the network’s requirements,overfitting will occur.Although we are in the era of big data now,certain special scenarios still make it difficult or impossible to obtain enough data for the network’s needs.Secondly,due to factors such as technical level and natural environment,directly obtained information cannot satisfy the network’s requirements for data and requires more detailed processing.The Retinex theory is based on the assumption that an image consists of reflection components and illumination components.Most current image enhancement methods are based on this theory;however,this hypothesis is relatively friendly towards weak light images but does not achieve ideal results for strong light images.To address these challenges,this paper proposes solutions as follows.To solve these problems mentioned above,this paper takes image data as input and uses network classification accuracy and heat maps as indicators to evaluate its effectiveness and reliability;designs color dimension representation functions to ensure color stability during image enhancement;designs an efficient fusion-based multi-dimensional image enhancement model to achieve high-quality enhancements while using peak signal-to-noise ratio(PSNR),structural similarity index(SSIM),mean absolute error(MAE)as evaluation criteria.Its main research contents include:(1)Proposing a folding fan-shaped data augmentation method which solves overfitting problem by achieving regularization effect through online multiple sample augmentation instead of simple cropping or splicing operations on images according to traditional Chinese folding fans’ styles while assigning labels according to different proportions occupied by each part in order to achieve more accurate classification results.Experimental results show that our proposed method can effectively prevent overfitting with 0.88% improvement in classification accuracy on two datasets respectively.(2)Proposing color dimension representation function B(x,y)to reduce color jitter during the enhancement process and maintain image color stability.Based on multi-scale image enhancement,our proposed method with color dimension representation can significantly improve the problem of color jitter in images,achieving optimal PSNR and MAE.(3)Proposing a multi-dimensional fusion-based enhancement network architecture to achieve efficient image enhancements for strong light images.The network architecture consists of three parts:illumination curve analysis,illumination attenuation loss function design,and illumination attenuation network construction.Our designed high-order version of the illumination curve is a three-dimensional matrix that satisfies the three-channel property requirement of an image by changing variables from general versions’ one-variable quadratic functions which are simple but sufficient for meeting networks’ needs.The illumination attenuation loss function design includes color consistency function,exposure restriction function,and lighting smoothness function to adjust the consistency of colors in images as well as their exposure values and lighting smoothness respectively.The illumination attenuation network is composed of nine convolutional layers with symmetric structures.Experimental results show that our proposed multi-dimensional fusion-based image enhancement network architecture can significantly improve the impact of strong light environments on images while improving their quality with optimal SSIM and MAE as well as second-best PSNR. |
