In this paper, based on muti-domain network (MDNet), fast deep learning for aircraft tracking (FDLAT) algorithm is proposed to track aircraft target. This algorithm uses feature-based transfer learning to make up the inferiority of small sample sets, uses specific data sets to update parameters of convolutional layers and fully connected layers, and use it to distinguish aircraft from background. After building the training model, we put the aircraft video sets into the model and tracked the aircraft using regression model and a simple line on-line update, to increase the speed while ensuring the accuracy. This algorithm achieves robust tracking for aircraft in rotation, similar targets, fuzzy targets, complex environment, scale transformation, target occlusion, morphological transformation and other complex states, and runs at a speed of 20.36 frames with the overlap reached 0.592 in the ILSVRC2015 detection sets of aircraft, basically meets the real-time application requirement of aircraft tracking.

Color transfer has been a hot research issue in the field of image processing and computer vision in recent years. The main purpose is to transfer the color of a target image to source image so that the source image has the same or similar color features with the target image. In practical applications for the color transfer of binocular stereoscopic images, the user may only need to transfer the color of the selected object while keeping the background color unchanged. For this purpose, a color transfer method based on the selected object is proposed in this paper. In the method, by assigning the object of the image by user, the accurate object is segmented via graph cut, and the probability density curves of color distribution between the selected object and the target image are matched to accomplish the color transfer. In order to enhance the viewing experience provided for the user, a non-linear disparity optimization is performed after the color transfer operation. According to the histogram feature of disparity map, the disparity mapping function is calculated, and the target disparity is obtained to enhance the depth sensation of the selected object. The experimental results demonstrate that the combination of stereoscopic color transfer and disparity remapping effectively enhances the stereoscopic viewing experience.

Multi-label image classification which is a generalization of the single-label image classification is aimed to assign multi-labels to the image to full express the specific visual concepts contained in the image. We propose a method based on convolutional neural networks, which combines attention mechanism and semantic relevance, to solve the multi label problem. Firstly, we use convolution neural network to extract features. Then, we apply the attention mechanism to obtain the correspondence between the label and channel of the feature map. Finally, we explore the channel-wise correlation which is essentially the semantic dependencies between labels by means of supervised learning. The experimental results show that the proposed method can exploit the dependencies between multiple tags to improve the performance of multi label image classification.

In view of the current state of the technology of the laser-guided weapon system that is vulnerable to fraudulent interference, a new idea using random sequence coding is proposed to improve its anti-jamming performance. By using the characteristics of better anti-interference performance of pseudo-random sequence, the laser active detection target system can not only achieve long-distance active target detection, but also effectively prevent external interference and improve the reliability of the system. The signal generation system is designed and implemented by combining Arduino IDE, Arduino UNO R3 microcontrollers, with oscilloscopes and YAG lasers, with good anti-interference performance. The system can be used for the study of new laser target indicators.

In order to highlight the texture details of the image while preserving the smooth region and saving the time to determine the fractional differential order, an improved adaptive fractional differential operator is proposed. Firstly, the classical Tiansi template is decomposed into four different directions, which are respectively convolved with the pixels to be processed to achieve the effect of enhancing the texture details of the image. Secondly, the current situation of the optimal differential order is determined by the experiment for the Tiansi operator. The local feature information of the image constructs a fractional order model with an adaptive ability, which can obtain more detailed information than the original image. The experimental results of multiple sets of different scene images show that the constructed adaptive fractional differential operators effectively enhance the texture details of the image. The subjective visual effects and objective evaluation indexes of the adaptive fractional differential operators are better than the original images. The average gradient, information entropy and contrast in the objective evaluation index are increased by 190.3%, 8.1%, and 18.3%, respectively. The average gradient and contrast are 45.0% and 9.6% higher than that of the Tiansi operator.

One of main challenges of driver assistance systems is to detect multi-occluded pedestrians in real-time in complicated scenes, to reduce the number of traffic accidents. In order to improve the accuracy and speed of detection system, we proposed a real-time multi-occluded pedestrian detection algorithm based on R-FCN. RoI Align layer was introduced to solve misalignments between the feature map and RoI of original images. A separable convolution was optimized to reduce the dimensions of position-sensitive score maps, to improve the detection speed. For occluded pedestrians, a multi-scale context algorithm is proposed, which adopt a local competition mechanism for adaptive context scale selection. For low visibility of the body occlusion, deformable RoI pooling layers were introduced to expand the pooled area of the body model. Finally, in order to reduce redundant information in the video sequence, Seq-NMS algorithm is used to replace traditional NMS algorithm. The experiments have shown that there is low detection error on the datasets Caltech and ETH, the accuracy of our algorithm is better than that of the detection algorithms in the sets, works particularly well with occluded pedestrians.

The output of fiber optic gyroscope (FOG) is easily affected by the temperature variations, so it leads to produce drift and the measurement accuracy of FOG is reduced. The traditional BP neural network is an optimization method of local search, which is easy to fall into local minimum, leading to the failure of network training. In order to optimize BP neural network, a temperature drift compensation method for FOG based on particle swarm optimization (PSO) and wavelet denoising is proposed. Firstly, the mechanism of FOG temperature drift is analyzed. Next, FOG static state test in different temperatures is finished. Finally, the FOG temperature drift model has been built by the method and compensate. The results show that the output standard deviation of FOG at different temperatures is reduced by 60.19%, and the compensation effect is better than traditional BP neural network.

The thermal infrared image of the human body directly reflects the temperature distribution of the human body surface. Based on in-depth analysis, the infrared image can provide intelligent diagnosis assistance for human diseases. This paper proposed two preprocessing algorithms, i.e., upper-lower body image-stitching and body image partitioning, for medical infrared image analysis. In the image stitching stage, the human body is first extracted from the background by local thresholding based on the characteristics of the actual imaging environment. Then the upper and lower body images are aligned and fused using binary and grayscale template matching. In the image partitioning stage, the key points of the part area are determined by the extremum-point analysis of the human contour. The human body is then partitioned into regions including head, trunk, limbs, etc. Experiments show that the proposed preprocessing algorithms produce satisfactory results in image-stitching and portioning, and can effectively support the quantitative and qualitative analysis of human body temperature distribution.

Small pixel targets in video images are difficult to detect. Aiming at the small pixel target in urban road video, this paper proposed a novel detection method named Road~~Net based on the YOLOv3 convolutional neural network. Firstly, based on the improved YOLOv3, a new convolutional neural network Road~~Net is designed. Secondly, for small pixel target detection depending on shallow level features, a detection method of 4 scales is adopted. Finally, combined with the improved M-Softer-NMS algorithm, it gets higher detection accuracy of the target in the image. In order to verify the effectiveness of the proposed algorithm, this paper collects and labels the data set named Road-garbage Dataset for small pixel target object detection on urban roads. The experimental results show that the algorithm can effectively detect objects such as paper scraps and stones, which are smaller pixel targets in the video relative to the road surface.

The target positioning algorithm of the traditional unmanned aerial vehicle (UAV) airborne optoelectronic platform introduces a large number of angle measurement errors, resulting in low target positioning accuracy. In this paper, a hybrid nonlinear algorithm of least squares and Gauss-Newton is proposed. Firstly, the Gauss-Newton iterative nonlinear target localization algorithm based on laser ranging value is derived. Then the rough solution of linear least square is used as the initial value of the nonlinear Newton iteration method for target location estimation. The algorithm combines the advantages of the simple and easy implementation of the least squares method and the high convergence accuracy of the Gauss-Newton method, and satisfies the requirements of the Gauss-Newton method for the initial value accuracy. Experimental results of measured data show that the longitude error of fixed target positioning results of this method is less than 1.37×10-5 degrees, the latitude error is less than 6.31×10-5 degrees, and the height error is less than 1.78 meters. And the processing time of each positioning is within 6 ms, which meets the requirements of real-time positioning.