机器视觉最小可察觉误差的建模及其优化OA

To improve the accuracy and reduce the complexity of the just noticeable difference(JND)model for machine vision,the framework design of existing JND models and the design of task accuracy loss,magnitude loss and spatial distribution loss in the model training constraints were investigated,and the improved machine vision JND model of Smooth-DMV-JND was proposed.The major components of the model were composed of three convolutional neural networks parts with saliency analysis module,edge extraction module and fusion analysis module,and the proposed model was trained with modified objective function by the designed two-stage training pipeline.The performance and complexity of the model were analyzed through experiments,and the application of Smooth-DMV-JND in image compression was presented.The results show that the proposed Smooth-DMV-JND model can provide more accurate estimation of JND for machine vision with shorter analysis time.Smoothed by the Smooth-DMV-JND model with maintaining about 88％classification accuracy,the processed images achieve bit-rate saving values of 17.68％for JPEG compression and 10.69％for BPG compression than that compressing the originals directly.The JND for machine vision can be modeled more effectively by Smooth-DMV-JND,which can guide the removal of redundancy in images and is beneficial to machine vision-oriented image compression.