Quantization And Detail Enhancement Method For High Dynamic Infrared Image

Infrared camera digital images generally have the characteristics of a wide dynamic range(14?16bits),a wide response range of infrared radiation energy,high temperature resolution,and are widely used in many fields such as military and civilian applications.However,most conventional display devices only support the display of 8-bit grayscale images(256-level grayscale),which needs to re-quantize the wide dynamic infrared image,thus inevitably leading to the dynamic range compression and gray level merging.Consequently,the image details are weakened or even lost.The key to the problem is simultaneously performing the enhancement and quantization.Additionally,Traditional methods such as linear quantization,decomposition-synthesis quantification,etc.rely on manual tuning of parameters,resulting in poor scene generalization.This research studies a wide dynamic range infrared image quantization and detail enhancement method based on deep convolutional neural networks.The main works are as follows:Firstly,the principle of image decomposition-synthesis quantization and detail enhancement methods based on bilateral filtering and guided filtering are analyzed.Combining evaluation indicators such as information entropy,detail enhancement factor EME,standard deviation,etc.,analyzing the effected law of algorithm parameters and their optimal settings,The results show that under the optimal parameter settings,the quantization method based on guided filtering achieves the best index performance.Secondly,to solve the problem of complicated parameter selection and poor scene adaptability based on the quantization and detail enhancement method based on guided filtering,We propose a method of infrared image quantization based on Dual CNN,constructing a detail layer sub-network Net D to learn image details,constructing a structure layer sub-network Net S to learn the large-scale structure of the image,and finally,adaptive fusion generates high-quality quantized images,According to our experiments the Dual CNN adaptively learns the guided filter quantization process under the optimal parameters,avoiding manual parameter tuning and improving scene adaptability.by big data learning.Finally,this research analyzes the promotion of wide dynamic range infrared image quantification and enhancement on subsequent target detection tasks.YOLOv2 is employed to make comparison of four quantization methods: linear gain control(AGC),platform histogram(PE),guided filtering,and our Dual CNN,as well as the original16-bit data respectively,Statistical analysis of large(Higher than 80 ×80),medium(20?80),and small(below 20 ×20)size target detection average accuracy(AP)and average recall(AR)performance,Experimentals results show that quantization and detail enhancement promote target detection performance;Optimal parameter-guided filtering quantization achieves the highest AP/AR performance,and the Dual CNN network quantization method achieves sub-optimal AP/AR performance.But it does not require manual parameter tuning,it is more practical.