Low-information Images Object Detection And Classification Based On Deep Learning

With the popularity of smartphones and vigorous promotion of monitoring equipment,photography has become an important information medium.However,due to the influence of the shooting environment,equipment,and subject,the sensor only receives a small amount or part of the optical signal.The image collected under this condition is called a low-information image.Low-information characteristics of the image make the image quality deteriorate or the whole area of the object cannot be detected.In addition,image classification,as the basis of high-level computer vision tasks,faces the problem of overfitting and low accuracy in the benchmark database.The topics of recovering a clear image from a low-information image,detecting the whole object,and improving the classification accuracy in the large sample benchmark database have important research value.In recent years,machine learning,especially deep learning,has developed rapidly.Compared with traditional algorithms,deep learning algorithms learn prior knowledge from big data,have stronger generalization capabilities,and can fit more complex mapping relationships between data.Benefitted by the above advantages,deep learning has been widely used in the field of computer vision and improves the test score of various computer vision tasks,but it still faces many challenges.In the face of extremely dark images with input light 100 times less than normal exposure,image restoration algorithms still have some disadvantages,such as low signal-to-noise ratio,unreal color and detail distortion,etc.After restoration,the test score of the object detection algorithm will reduce,even the detection will fail in the face of the objects with an invisible part of the object.Multilayer neural networks have the problems of gradient attenuation,over-fitting,model degradation,and low test accuracy for large sample image classification tasks.In response to the above challenges,this dissertation focuses on lowinformation image restoration based on autoencoders,incomplete object detection,and multibranch random shake deep layer image classification based on the Residual module.The innovations and contributions of this dissertation are summarized as follows:1)Supervised end-to-end low-information dark image restoration algorithm based on AutoencoderThere are many types of low-information images according to the collection environment and methods,two of them are studied in this dissertation.One kind of low-information image is the image collected in an extremely dark environment in the air,and another is the underwater image captured in a dark underwater environment with suspended particles.Aiming at the first kind of low-information image,an end-to-end supervised image restoration algorithm based on small samples data aligned pixel by pixel is proposed in this dissertation.The training data of this algorithm is 20 high-quality image pairs of "extremely dark image-clear image",which are trained by the autoencoder network based on the U-Net network.The algorithm can restore the100 times reduced exposure images collected on 1 lumen low light environment.The proposed NL2 LL method is used to collect the paired images,that is,dark images are collected by adjusting camera parameters under a normal light environment.NL2 LL improves the acquisition conditions of images,increases the possibility of image alignment at the pixel level,and provides a new solution for image acquisition in a harsh environment.Aiming at the second kind of low-information images,the mixed scattering model in atmosphere and water named MSM-AW is built,which proves the feasibility of placing the subject outside the tank,based on which the virtual underwater images are collected.An underwater image restoration model based on an autoencoder network is proposed.Data is input into an underwater image restoration network based on Autoencoder,the test results show that the proposed algorithm surpasses traditional algorithms and most machine learning algorithms.2)Object detection DIDA algorithm of low-information incomplete targetBecause of the special shooting environment such as object occlusion,self-occlusion,color texture blending into the background,part area of the target is not visible.In this case,the object detection algorithm is difficult to detect the hidden area of the target,and the detection efficiency decreases or even fails.For solving this difficult,an algorithm named DIDA is proposed that uses object detections algorithm twice combing Generative Adversarial Networks.DIDA first performs object detection once,then forcibly generates the missing parts,and identifies the hidden part of the object by using the object detection algorithm again.In the experiment,a virtual image of an object with a part of the missing area is generated,and the validity of the DIDA algorithm is verified on the virtual image.This algorithm solves the problem that the single object detection algorithm cannot detect the incomplete target correctly,and has strong practical value.3)Multi-branch randomly shake super deep anti-overfitting image classification network based on residual networkThe proposed MBRMDAC-RES algorithm adds a branch to the classic Res Net network,calculates the product of the data and the random multiplier at each branch to augment the data in the network by data randomness.The network sets the constraint that the sum of two random multipliers is 1 to control the sum of data passing through the two branches.Furthermore,an adaptive learning rate method is proposed to speed up the fitting process.Cosine decay is performed firstly,then the adaptively corrected method based on the test error is used during the training stage.Appropriate batch size is chosen to balance speed and accuracy.The Top-1error rate of this algorithm on the standard database CIFAR-10 exceeds that of the classic Res Net network.The algorithm verifies that expanding the width and depth of the network can improve the performance of the network,and has strong capabilities of anti-overfitting and weakening gradient disappearance.