Research And System Implementation Of Pixel-level Image Exposure Evaluation Method

Image exposure evaluation is an important part of image aesthetics evaluation,which focuses on analyzing whether the brightness and darkness of various parts of the image are appropriately displayed and evaluating the overall light coordination effect of the image.It essentially quantifies human aesthetic emotional responses to the brightness and darkness stimuli in images.With the improvement of people’s living standards,the demand for improving image aesthetics has increased.Exposure,as one of the three elements of photography,is a key factor that affects image aesthetics.However,defining and quantifying image exposure is still a major challenge.Traditional image exposure evaluation usually defines features related to brightness and darkness in images based on prior knowledge from a few experts and then uses statistical analysis and threshold setting to make judgments.However,exposure is a subjective aesthetic issue,and traditional methods that use a single feature have poor generalization ability and inaccurate evaluation results.In recent years,there have been methods based on deep learning,but these researchers have mostly considered only the overall exposure level or visual effect of the image and have not conducted specific analysis and evaluation of the exposure problem in images.To address these issues,this paper models the classical photography theory of exposure and divides the human evaluation process of image exposure into three key steps:first,perception of the brightness and darkness of each area of the image;second,judgment of the importance of each object in the image;and finally,appreciation of the overall exposure effect of the image.Specifically,the contributions of this paper are as follows:1)This paper proposes a two-stage deep learning-based exposure evaluation model.To address the problem that existing methods cannot evaluate the differential distribution of image exposure well,we design an encoder-decoder network structure in the first stage of the model to predict the exposure level of each pixel in the image.At the same time,to fully utilize the neural network’s perception ability for images,we design a double-decoder structure in the first-stage model to perceive the importance of each area of the image.In the second stage of the model,we explore different feature fusion schemes to more accurately evaluate the overall exposure effect of the image by fusing pixel-level exposure problem information and information on the importance of each area of the image.2)This paper constructs a dataset for image exposure evaluation.Previous datasets for exposure evaluation had shortcomings in data scale and annotation quality,which were insufficient to drive the training and development of exposure evaluation algorithms.This paper constructs a dataset IEA-40k specifically for image exposure evaluation.In data collection,IEA-40k collects different images from various devices and diverse scenes,and in data annotation,it adopts a quality control method based on anchor points to ensure the annotation quality of the dataset.The proposal of this dataset provides strong support for research in the field of image exposure and contributes to its further development.3)This paper proposes an image exposure evaluation system.Based on the image exposure evaluation model of the above research,an image exposure evaluation system is developed in this paper.The functions of the system mainly include the assessment of the pixel level exposure of the image,the assessment of the overall exposure effect of the image and the function of personalized photo album image ranking.This system builds a set of distributed real-time reasoning system based on Apache Kafka framework.Through load balancing strategy,A series of optimizations including micro-batch processing strategy,edge computing shunt,dynamic batch design,semi-precision FP16 reasoning strategy,Tensor RT reasoning acceleration and Tensor RT quantization compression processing are necessary to optimize the real-time response capability of the system.