This study proposed a novel intensity weighting approach using a convolutional neural network (CNN) for fast and accurate optic disc (OD) segmentation in a fundus image. The proposed method mainly consisted of three steps involving CNN-based importance calculation of pixel, image reconstruction, and OD segmentation. In the first step, the CNN model composed of four convolution and pooling layers was designed and trained. Then, the heat map was generated by applying a gradient-weighted class activation map algorithm to the final convolution layer of the model. In the next step, each of the pixels on the image was assigned a weight based on the previously obtained heat map. In addition, the retinal vessel that may interfere with OD segmentation was detected and substituted based on the nearest neighbor pixels. Finally, the OD region was segmented using Otsu’s method. As a result, the proposed method achieved a high segmentation accuracy of 98.61%, which was improved about 4.61% than the result without the weight assignment.

Headlight is the most explicit and stable image feature in nighttime scenes. This study proposes a headlight detection and pairing algorithm that adapts to numerous scenes to achieve accurate vehicle detection in the nighttime. This algorithm improved the conventional histogram equalization by using the difference before and after the equalization to suppress the ground reflection and noise. Then, headlight detection was completed based on this difference as a feature. In addition, the authors combined coordinate information, moving distance, symmetry, and stable time to implement headlight pairing, thus enabling vehicle detection in the nighttime. This study effectively overcame complex scenes such as high-speed movement, multi-headlight, and rains. Finally, the algorithm was verified by videos of highway scenes; the detection rate was as high as 96.67%. It can be implemented on the Raspberry Pi embedded platform, and its execution speed can reach 25 frames per second.

Fingerprint quality assessments are generally used to evaluate the quality of images obtained from fingerprint sensors, and effective fingerprint quality assessment methods are crucial to establishing high-performance biometric identification systems. The use of fingerprint quality assessments helps improve the accuracy of fingerprint registration and user satisfaction. NIST Fingerprint Image Quality (NFIQ) is a popular fingerprint quality assessment algorithm; however, it is unable to provide high-quality assessments for some partial fingerprint images obtained from mobile device sensors. In this study, a hybrid fingerprint assessment framework that integrated texture and geometric features was examined. The final quality assessment values obtained by the framework were higher than those obtained using NFIQ, effectively elevating the performance of existing NFIQ algorithms and expanding its scope of application for different fingerprint images.

With rapid developments in techniques related to the internet of things, smart service applications such as voice-command-based speech recognition and smart care applications such as context-aware-based emotion recognition will gain much attention and potentially be a requirement in smart home or office environments. In such intelligence applications, identity recognition of the specific member in indoor spaces will be a crucial issue. In this study, a combined audio-visual identity recognition approach was developed. In this approach, visual information obtained from face detection was incorporated into acoustic Gaussian likelihood calculations for constructing speaker classification trees to significantly enhance the Gaussian mixture model (GMM)-based speaker recognition method. This study considered the privacy of the monitored person and reduced the degree of surveillance. Moreover, the popular Kinect sensor device containing a microphone array was adopted to obtain acoustic voice data from the person. The proposed audio-visual identity recognition approach deploys only two cameras in a specific indoor space for conveniently performing face detection and quickly determining the total number of people in the specific space. Such information pertaining to the number of people in the indoor space obtained using face detection was utilized to effectively regulate the accurate GMM speaker classification tree design. Two face-detection-regulated speaker classification tree schemes are presented for the GMM speaker recognition method in this study—the binary speaker classification tree (GMM-BT) and the non-binary speaker classification tree (GMM-NBT). The proposed GMM-BT and GMM-NBT methods achieve excellent identity recognition rates of 84.28% and 83%, respectively; both values are higher than the rate of the conventional GMM approach (80.5%). Moreover, as the extremely complex calculations of face recognition in general audio-visual speaker recognition tasks are not required, the proposed approach is rapid and efficient with only a slight increment of 0.051 s in the average recognition time.

In this article, we present a U-Net convolutional network for solving insufficient data problems of color patches in colorimetric characterization. The U-Net network uses data augmentation annotated over 6,885,222 colors, 32,027,200 color patches, and 2,098 billion pixels directly from only eight standard colorimetric images of ISO 12640 (CIELAB/SCID). By applying the U-Net network trained on big augmented data, the pixel-wise colorimetric characterization is implemented from digitalized red, green, blue image samples to ISO 12640 (CIELAB/SCID) CIELAB standard colorimetric images. The performance efficiency of the U-Net network is superior to that of the convolutional neural network on both training and validating epochs. Moreover, pixel-wise color colorimetric characterization is achieved using the intelligent machine vision of U-Net integrated with a data augmentation technique to overcome the drawback of complex color patches and labor-intensive tasks. This study might improve colorimetric characterization technology with a resolution of 2560-by-2048 for over 4 million pixels. The results reveal that U-net with pixel-wise regression enhances the precise colors of images, taking detail and realism to a new level.

In recent years, a variety of mobile road measurement equipment has emerged and become an important means of collecting spatial information. As an important part of the mobile road measurement system, a camera’s function implementation and data accuracy largely depend on its internal parameters and the rotation and translation parameters corresponding to the world coordinate system. Based on this and on the traditional camera calibration method, radial and tangential distortion for monocular camera calibration is introduced in this article to establish a calibration model, and the nonlinear least-squares Levenberg–Marquardt optimization algorithm is used in iterative calculation. The parameters provide a solution to the problem of rapid calibration of camera modules in mobile road measurement systems. The camera parameters obtained by the calibration algorithm in this study are used for visual reconstruction. Compared with two Zhang Zhengyou calibration methods optimized by the Gauss–Newton method, the former has an average pixel offset of 0.28 pixel and the latter has deviations of 0.66 and 0.38 pixel. Using a monocular camera to collect data on geometric targets on a road, the average relative error does not exceed 2.16%. Experiments show that this method can obtain calibration results quickly and accurately.

Yongding River is one of the five major river systems in Beijing. It is located to the west of Beijing. It has influenced culture along its basin. The river supports both rural and urban areas. Furthermore, it influences economic development, water conservation, and the natural environment. However, during the past few decades, due to the combined effect of increasing population and economic activities, a series of changes have led to problems such as the reduction in water volume and the exposure of the riverbed. In this study, remote sensing images were used to derive land cover maps and compare spatiotemporal changes during the past 40 years. As a result, the following data were found: forest changed least; cropland area increased to a large extent; bareland area was reduced by a maximum of 63%; surface water area in the study area was lower from 1989 to 1999 because of the excessive use of water in human activities, but it increased by 92% from 2010 to 2018 as awareness about protecting the environment arose; there was a small increase in the built-up area, but this was more planned. These results reveal that water conservancy construction, agroforestry activities, and increasing urbanization have a great impact on the surrounding environment of the Yongding River (Beijing section). This study discusses in detail how the current situation can be attributed to of human activities, policies, economic development, and ecological conservation Furthermore, it suggests improvement by strengthening the governance of the riverbed and the riverside. These results and discussion can be a reference and provide decision support for the management of southwest Beijing or similar river basins in peri-urban areas.

In order to reconstruct and identify three-dimensional (3D) images, an image identification algorithm based on a deep learning compensation transformation matrix of main component feature dimensionality reduction is proposed, including line matching with point matching as the base, 3D reconstruction of point and line integration, parallelization automatic differentiation applied to bundle adjustment, parallelization positive definite matrix system solution applied to bundle adjustment, and an improved classifier based on a deep compensation transformation matrix. Based on the INRIA database, the performance and reconstruction effect of the algorithm are verified. The accuracy rate and success rate are compared with L1APG, VTD, CT, MT, etc. The results show that random transformation and re-sampling of samples during training can improve the performance of the classifier prediction algorithm under the condition that the training time is short. The reconstructed image obtained by the algorithm described in this study has a low correlation with the original image, with high number of pixels change rate (NPCR) and unified average changing intensity (UACI) values and low peak signal to noise ratio (PSNR) values. Image reconstruction effect is better with image capacity advantage. Compared with other algorithms, the proposed algorithm has certain advantages in accuracy and success rate with stable performance and good robustness. Therefore, it can be concluded that image recognition based on the dimension reduction of principal component features provides good recognition effect, which is of guiding significance for research in the image recognition field.

Object detection and tracking is an indispensable module in airborne optoelectronic equipment, and its detection and tracking performance is directly related to the accuracy of object perception. Recently, the improved Siamese network tracking algorithm has achieved excellent results on various challenging data sets. However, most of the improved algorithms use local fixed search strategies, which cannot update the template. In addition, the template will introduce background interference, which will lead to tracking drift and eventually cause tracking failure. In order to solve these problems, this article proposes an improved fully connected Siamese tracking algorithm combined with object contour extraction and object detection, which uses the contour template of the object instead of the bounding-box template to reduce the background clutter interference. First, the contour detection network automatically obtains the closed contour information of the object and uses the flood-filling clustering algorithm to obtain the contour template. Then, the contour template and the search area are fed into the improved Siamese network to obtain the optimal tracking score value and adaptively update the contour template. If the object is fully obscured or lost, the YoLo v3 network is used to search the object in the entire field of view to achieve stable tracking throughout the process. A large number of qualitative and quantitative simulation results on benchmark test data set and the flying data set show that the improved model can not only improve the object tracking performance under complex backgrounds, but also improve the response time of airborne systems, which has high engineering application value.