Weakly illuminated image enhancement as a kind of pre-processing technologies is widely used in various computer vision tasks. Traditional image enhancement methods has poor robustness. While, methods based on existing convolutional neural networks (CNNs) estimate the enhanced image from weakly illuminated image directly without injecting the visual attention mechanism, ignoring weakly illuminated regions and leading to inaccuracy result. To resolve this problem, we propose an attentive residual dense network for weakly illuminated image enhancement. The proposed network contains two parts: attentive recurrent network and residual dense network. With the guidance of the illumination map, attentive recurrent network focuses more and more on the weakly illuminated regions and generates the attentive illumination map following a coarse-to-fine strategy via the recurrent architecture. This attentive illumination map concatenated with the weakly illuminated image are injected into the subsequent residual dense network to ensure this network assign more computational source to weakly illuminated regions and estimate enhanced image accurately. The experiments demonstrate that our method achieves favorable performance against that of existing image enhancement methods based on synthetic images and real images.

To solve the blurring, low contrast and color distortion problem of underwater image caused by light absorption and scattering effects in the underwater environment, an underwater image enhancement algorithm based on the Inception-Residual and generative adversarial network is proposed. Firstly, the degraded underwater image is scaled to a size of 256×256×3 to obtain a data set for the training model. The Inception module, residual idea, encoding and decoding structure and generative adversarial network are combined to build an IRGAN(Generative Adversarial Network with Inception-Residual) model to enhance underwater images. Then, a multi-loss function including global similarity, content perception and color perception is constructed to constrain the antagonistic training of generative network and discriminant network. Finally, the degraded underwater image is processed by the trained model to obtain a clear underwater image. The experimental results show that, compared with the existing enhancement methods, the average values of the PSNR, UIQM and IE indicators of the underwater images enhanced by the proposed algorithm are improved by 13.6%, 4.1% and 0.9%, respectively, compared with the second place. In subjective perception and objective evaluation, the sharpness, contrast enhancement and color correction of the enhanced underwater image are improved.

In order to solve the problems of texture details loss, indistinguishable treatment of all channel and spatial feature information and poor deblurring effect in the process of image restoration, an image deblurring method based on dual task convolutional neural network is proposed. The image deblurring task is divided into deblurring sub task and high frequency detail restoration sub task. Firstly, a coding and decoding sub network model based on Residual Attention Module and Octave Convolution Residual Block is proposed, which is used in image deblurring sub task. Secondly, a high frequency detail recovery sub network model based on Double Residual Connection is proposed, which is used in high frequency detail recovery sub task. The two subnetworks are combined in parallel, and the average absolute error loss and structure loss are used to constrain the training direction to achieve image deblurring. The experimental results show that the proposed method has strong image restoration ability and rich detail texture, the peak signal-to-noise ratio (PSNR) is 32.427 6 dB, and the structure similarity(SSIM) is 0.947 0. Compared with the current advanced deblurring algorithm, it can effectively improve the image deblurring effect.

Aiming at the problems of huge model and multi-scale object segmentation in the classical semantic segmentation algorithm, an efficient multi-scale image semantic segmentation method based on ASPP and HRNet is proposed. Firstly, the basic block is improved by using deep separable convolution combined with 1 * 1 convolution to reduce the model parameters. Secondly, a batch normalization layer (BN) is added after all convolution layers and before the relu activation function to improve the dead relu problem. Finally, the improved ASPP module based on the hybrid dilated convolution is added, and the advantages of the two are fused by using the parallel upsampling channels to obtain the spatial accurate segmentation results. The RE-ASPP-HRNet is proposed. Results on Pascal voc2012 and CityScapes show that the proposed method is effective. Compared with the original HRNet, it can improve the accuracy of 0.8% or 0.5% MIoU, and reduce the number of parameters by 1/2 and memory by 1/3. We implement a more efficient and reliable multi-scale semantic segmentation algorithm.

Optical remote sensing images are affected by the complexity of the background and the large amount of semantic information, and there are still certain deficiencies in detection accuracy and efficiency. This paper proposes an SSD300 network model based on Resnet50 for feature extraction, adding the attention mechanism CBAM module and feature fusion FPN module, and adopting the Soft-NMS strategy to select the final prediction frame to detect aircraft targets in remote sensing images more effectively. Finally, training on 2 150 aircraft remote sensing image data sets, when the IoU is 0.5 and 0.75, the average accuracy MAP reaches 92.54% and 63.44%, which are 5.04% and 11.38% higher than the previous algorithm model, and the detection speed reaches 13.4 FPS. Experimental results show that this method can effectively improve the detection ability of objects and quickly and accurately detect aircraft objects in the airport area, effectively reducing the missed detection rate of aircraft objects and improving detection accuracy and speed.

Aiming at the problems that the deep features are difficult to extract, the real-time performance cannot meet the requirements, and the bounding box positioning is not accurate enough of the real-time object detection algorithm Tiny YOLOv3, an improved lightweight detection model MTYOLO (MdFPN Tiny YOLOv3) is proposed. The model constructs multi-directional feature pyramid network (MdFPN) instead of simple concatenation to adequately complete the extraction and fusion of multi-layer semantic information. The deep separable convolution is used instead of standard convolution to effectively reduce the complexity of the network and improve the real-time performance of detection. The complete IOU loss (CIOU loss) is used instead of MSE as the regression loss function, which greatly improves the regression accuracy of bounding box. The results which MTYOLO is tested on PASCAL VOC and COCO datasets show that the mAP of the improved model can reach 78.7% and the detection speed can reach 205 fps.

During the period of 2019-nCoV controlling, to prevent the spread of the virus, it is necessary to regulate the coverage of mask wearing in densely populated places such as airports and stations. In order to effectively monitor the coverage of mask wearing of crowd, this paper proposes a lightweight mask detection algorithm based on improved YOLOv4-tiny. Following the backbone network of YOLOv4-tiny, a spatial pyramid pooling structure is introduced to pool and fuse the input features at multi-scale, which makes the receptive field of the network enhanced. Then, combined with the path aggregation network, multi-scale features are fused and enhanced repeatedly in two paths to improve the expressive ability of feature maps. Finally, label smoothing is utilized to optimize the loss function for modifying the over-fitting problem in the training process. The experimental results show that the proposed algorithm achieves 94.7% AP and 85.7% AP on mask target and face target respectively (at real-time speed of 76.8 FPS on GeForce GTX 1050ti), which is 4.3% and 7.1% higher than that of YOLOv4-tiny. The proposed algorithm meets the accuracy and real-time requirements of mask detection tasks in various scenes.

The difficulty of 3D target detection in practical engineering applications lies in the high price of depth perception equipment, poor point cloud quality, lack of rich texture information, difficulty in creating 3D data training sets. This paper proposes a three-dimensional target position estimation method based on instance segmentation. It can be used in a variety of sensors, such as camera-radar, binocular camera, etc. Firstly, the target segmentation is performed under the two-dimensional image, the targets depth image are extracted and RGB image according to the target segmentation mask, and it is converted into a rough point cloud. Finally, the abnormal noise points is removed to obtain a fine target point cloud. Tested on the KITTI data set, the AP can reach 50%. The results show that this method can accurately estimate the target location information. The method proposed in this paper does not need 3D data training set can quickly and accurately extract the point cloud of three-dimensional objects, and only use a two-dimensional detector to achieve the purpose of three-dimensional object detection.

In order to realize the non-destructive inspection of the internal defects of composite materials and the accurate marking of the defect positions, this paper proposes a projection labeling method for infrared thermal wave detection. This system uses thermal imaging camera to obtain the infrared image sequence of the sample under the flashlights pulse excitation, and processes it by the pulse phase thermography optimized by sampling time to enhance the detection effect of the defect, and then uses the automatic threshold to extract the defect position and project the extraction results onto the sample surface through a projector. To solve the problem of image distortion caused by factors such as the viewing angle difference between the projector and the thermal imager, this paper proposes a visible-infrared camera-assisted calibration method. The auxiliary camera is used to calibrate the projector and establish the spatial coordinate relationship between the projector and the thermal imager to correct the distortion of the defect extraction results and improve the accuracy of defect marking. The measurement experiment results show that the defect non-destructive testing and marking method can mark the area of different size defects within 10%, and the center mark errors are within 3 mm, which can accurately mark the internal defects of the material.

The miscellaneity of grid information in the existing target tracking algorithm is high, which increases the complexity of the tracking algorithm and makes the calculation amount of tracking too large, which affects the tracking effect. Therefore, the primary goal is reducing the complexity of the tracking process. In this paper, under the condition of considering the feature constraints, the optical flow field model is used to achieve the effective tracking of multi-channel video targets. The video random variable is set to control the variation of the numerical model of optical flow field, so as to control the deviation and extract the normalized features of multi-channel video, and the global pooling processing method is introduced to solve the problem of complex calculation of multi-dimensional features in the traditional method. Firstly, according to the numerical relationship between image pixel intensity and time parameters, the deviation of the mean value of optical flow field model is controlled to obtain video features. Then, the activation function is used to map the multi-channel parameters into fixed parameters, and the multi-channel feature constraint relationship is constructed to control the increasing dimension of the feature parameters and reduce the calculation dimension of the data. Finally, the background and target areas in the video image are divided, and the video target is tracked by constructing the pixel motion model. After setting the initial parameters of the experimental instrument, the video clips are randomly selected from the database as the experimental object, and the image resolution of different frames is set to simulate them as multi-channel video information. The performance tests of the traditional algorithm and the proposed algorithm are carried out, and the results show that the algorithm in this paper can not only guarantee the tracking accuracy higher than 0.90, but also control the successful tracking rate above 96%, and the tracking process takes only 5.5 s at most. The algorithm in this paper can track multi-channel video target with low computational complexity and has low computational cost.

With the spread of new coronary pneumonia, in order to accurately diagnose COVID-19, this article proposes an improved new coronary pneumonia recognition algorithm based on convolutional neural network, namely the ARS-CNN algorithm. Based on the CNN network structure, this algorithm adds new functional modules. Firstly, in order to capture the multi-scale feature information of different receptive fields and strengthen the networks use of image features, a jump connection RFB structure is proposed. Secondly, the problem of local information loss caused by the reduction of image resolution during the feature extraction process of the network is improved by short-connecting the aspp module. Finally, the attention mechanism GC module and the sSE module are merged to achieve the screening of feature information and the interaction between feature information, thereby improving the accuracy of new coronary pneumonia recognition. Experiments on the public COVID-19 Chest X-ray Database data set show that the weighted average accuracy, precision, recall, FI score, and specificity of the algorithm proposed in this article are 98.22%, 97.91%, 9795%, 97.92%, 98.33%, respectively. Compared with other classification algorithms, the algorithm proposed in this paper can efficiently recognize lung diseases and has higher recognition performance.

In order to solve the problems of occlusion, large differences in styles between domains and cameras in the research of unsupervised person re-recognition, this paper proposes an unsupervised domain adaptive model based on deformable convolution. Aiming at the occlusion problem in the feature extraction process, a CNN model based on deformable convolution is proposed. In the pre-training stage, it is proposed to apply SPGAN to directly reduce the difference between domains. During the training process, it is proposed to use CycleGAN to generate images of different camera styles to alleviate the problem of camera style differences. A multi-loss collaborative training method is proposed to realize the iterative optimization of CycleGAN and re-used CNN models to further improve the recognition accuracy. The experimental results show that the method proposed in this paper is tested in the source domain DukeMTMC-reID/Market-1501 and the target domain Market-1501/DukeMTMC-reID, and mAP and Rank-1 reach 68.7%, 64.1% and 88.2%, 78.1%, respectively. The model proposed in this paper effectively alleviates the problems of pedestrians being occluded, and large differences in styles between domains and cameras. Compared with the existing methods, it has a better recognition effect.

At present, a large number of offline videos are stored in the servers of the surveillance network in criminal investigation departments of public security organs, and the face retrieval system is designed. The new Quadruplet Network converges faster than the familiar networks such as Alexnet, Googlenet, VGGNet and ResNet. Because of the shared weight design of the network, the retrieval has a high precision, Average Retrieval Precision(ARP) and model training accuracy, and the system has good robustness. The image depth features can be shared quickly online between the cameras. The proposed method is effective, with ARP of 98.74% and a model training accuracy of 99.51%, and a frame rate of 28 FPS.

In order to extend the depth of field and improve the reliability of the system without reducing the resolution of the imaging optical system for space targe, this paper proposes and designs an imaging optical system for space target based on wavefront coding technology to extend the depth of field. On the basis of the initial imaging optical system for space target, the cubic phase plate is optimized based on the consistency of the modulation transfer function curves of different object distances and the recoverability of the images. The FOPD algorithm is used to achieve a better intermediate images recovery effect. The effects of the initial optical system and the wavefront coding optical system are compared, and the results show that the depth of focus and depth of field of the wavefront coding optical system have been significantly expanded. The total length of the optical system is 41 mm, the focal length is 20 mm, the working wavelength is 850 nm. The simulation experiment results show that without the need for focusing mechanism, the depth of focus of the system is 28 times that of the initial system, and the depth of field of the system is extended from 0.956~1.100 m to 0.5~130 m. The system can meet the needs of large depth of field imaging without focusing mechanism in the fields of space rendezvous and docking, satellite capture and so on.