Research And Application Of Weak Feature Visual Perception Algorithm Based On Deep Learning

With the development of artificial intelligence,computer vision technology has becomes one of the important research topics.In the face of the rapid growth of image data,the scope of application of visualization based on feature perception is expanding,and how to perceive the target characteristics under different backgrounds and environments effectively has become a hotspot in the field of computer vision.The method based on target feature perception is widely used in the visual tasks of detection,recognition and segmentation,and the task is seriously constrained by the characteristic representation.Especially for the environment weakening the target characteristics,the target is similar with the background or the target is occluded,the feature is weakened,and the performance of the target information is decreased,the extracted features can not describe the target.Therefore,in order to solve the problem of weak feature visual perception,this paper puts forward the corresponding solutions for the three forms of weak features with deep learning.This paper mainly studies from the following three aspects(9):1)Aiming at the problem of haze weakening the scene features,the paper propose a feature-aware enhancement algorithm to the problem caused by haze weaken target features in this paper,it is based on the generative adversarial dehaze nets for the problem of color shift and the priori limitation.Firstly,the haze feature is extracted from the image.Secondly,the initial extraction feature is converted to the medium transmission by the generate network.Finally,dehaze is in according to the multiplelight scattering model.The results of the proposed algorithm are compared with the state-of-the-art algorithm,and the dehaze result is more realistic with the real scene.It can save more details and introduce less interference.2)Aiming at the problem of the similarity between the background elements and the target,a feature detection algorithm based on region proposal network(RPN)is proposed for the camouflage color target in this paper.Firstly,the image is subjected to color enhancement to amplify the difference between the target.Secondly,the target detection is performed on the enhanced results,and the target is detected by faster region convolutional neural network(Faster R-CNN).Compared with the the state-of-the-art target detection algorithm,the proposed algorithm has higher accuracy,and can more effectively identify the target in the interference background.It also reduces the leakage detection problem.3)Aiming at the problem of the detecting features missing caused by occlusion,the occlusion target detection algorithm based on cooperative sensing deep neural network is proposed to solve the problem of occlusion target detection in this paper.Firstly,the Faster R-CNN is used to detect the parents and the daughters of the target in the image.Secondly,the non-maximum suppression(NMS)process is performed on the candidate region obtained by the the daughters detection.Finally,the daughters is used to coordinate the fusion to realize the correction of the parents.Compared with the the state-of-the-art target detection algorithm,the detection performance shows the higher accuracy and fewer phenomenons of non-detection.