Research And Implementation Of Low Illumination Image Enhancement Model Compress Algorithm For Embedded Devices

Current high-level computer vision tasks based on deep learning mostly rely on high-quality image and video resources.Images collected in low-illuminance scenes are affected by external environmental factors,and their imaging quality is often very poor and cannot be directly applied to high-level vision tasks.In recent years,researchers have proposed related image enhancement models to try to recover the semantic information of low-illuminance images,thereby improving the accuracy of related visual tasks in low-illuminance scenes,while also satisfying people's photography needs in different environments.Although there are related algorithms that can meet the task requirements of low-illuminance image enhancement,due to their huge amount of parameters and computing power requirements,as well as the limitations of the computing power of mobile devices,they cannot be used in actual production and life.In order to overcome the problem of image enhancement model deployment on the end-side,we propose an image enhancement model compression algorithm for embedded devices.Specifically,we propose a feature extraction enhancement module,which can be embedded in any convolutional neural network to improve the accuracy of the algorithm without introducing parameter and computing power requirements.In addition,the deep separable convolution structure and channel-scaling-factors are introduced to construct lightweight models of different sizes.We propose knowledge distillation architecture and parameter quanti-fication methods for generation adversarial network.Under the premise of loss of certain performance,the model parameters and computing power requirements are greatly reduced.Finally,based on the compressed lightweight model,we develop a low illumination image enhancement prototype system combined with embedded devices,which can be applied in actual production and life.In order to evaluate the effectiveness of the low-illuminance image enhancement model compression algorithm proposed in this article for embedded devices,we conduct sufficient experimental verifications.The results show that our enhanced model compression algorithm has better performance than the benchmark model and is more friendly to the deployment.The image enhancement prototype system based on the lightweight model can be processed in real time on the end-side and has good efficiency.