Embedded Image Processing System Based On Low Illumination CMOS

The demand for low illumination imaging capability and localization of man-portable optoelectronic equipment is rapidly increasing in domestic markets.This paper focuses on exploring the key technologies of domestic embedded system and low illumination image enhancement.Additionally,it presents the designs of a low illumination CMOS embedded image processing system based on domestic low illumination CMOS and the Heath SD3403 platform.First,this paper analyzes the technical specifications and functional requirements of the system,and composes the overall scheme for the low-illumination CMOS-based embedded image processing system through a literature review.The system is divided into two main parts:the hardware module and the software module.Subsequently,the workflow of the overall system software is developed.Secondly,this paper completes the hardware selection and circuit design of the embedded system for the hardware module,in accordance with the overall system design plan.The selection of detector and processor chips is based on the requirements of localization.The hardware circuit design encompasses the minimum system and peripheral circuit components,ultimately leading to the completion of the embedded system board.Then,the design of the image acquisition module is realized using the Hessian MPP platform,while the image enhancement module is designed based on the Open CV and SVP platforms for the software module.The image enhancement module,which is based on the Retinex Net network,addresses the defects of the original network,such as color distortion,blurred edge information and limited embedded portability of the two-segment network.Improvements are made by replacing the Decom Net part of the network with an adaptive image decomposition algorithm based on flat exponents to obtain incident and reflected components.This solution resolves the challenges associated with embedding and porting Retinex Net.Additionally,the attention mechanism and color compensation function are introduced into the Enhance Net part of the network to rectify the color distortion problem of Retinex Net.The reflected components,containing most of the noise and edge information,are denoised and sharpened to solve the problem of blurred edge information after enhancement.Furthermore,considering that SD3403 only supports the Caffe model,this paper successfully ports the improved neural network algorithm and generates the corresponding WK executable file.Finally,experimental tests are conducted in a controlled dark room with varying levels of low illumination,enabling image acquisition.The results are objectively compared and evaluated against a control system,demonstrating that the system designed in this paper can produce clear imaging even at 0.01 lux illumination.The overall performance of the system is substantially improved compared to the imaging results of the control system.Additionally,the system is utilized to capture night scenes of a building at a distance of 400 m and night scenes without lights at a distance of 100 m in outdoor environments,resulting in high-quality images.These outcomes substantiate the practical applicability of the proposed system in this paper and highlight its adaptive characteristics,making it suitable for image enhancement in various complex environments.