Real-time object detection remains an important topic in computer vision. Balancing the accuracy and speed of object detectors is a formidable challenge for both academic researchers and industry practitioners. In this paper, considering the latest models may be somewhat over-optimized for anchor-free pipes, we elect to use YOLOX as our baseline and introduce a series of enhancements, forming in a new high-performance detector named YOLOAX. To further exploit the power of the attention mechanism, we devise multi-dimensional attention-based modules which can activate CNNs, emphasizing regions of interest and boosting the capacity to learn the informative representations from feature maps. Moreover, we introduce a new label assignment strategy called STA, along with a novel loss function named GEIOU Loss, to further refine our object detector’s performance. Extensive ablation studies on the COCO and PASCAL VOC 2012 datasets are provided to validate our proposed methods. Our YOLOAX series is trained solely on the COCO dataset from scratch, without any prior knowledge, surpassing YOLOX series by a margin of 4.0% AP. Especially, YOLOAX-X achieves an impressive 55.2% AP on the COCO 2017 test set while maintaining a real-time speed of 82.4 fps.

Few-shot object detection (FSOD) aims to address the challenge of requiring a substantial number of annotations for training in conventional object detection, which is very labor-intensive. However, the existing few-shot methods achieve high precision with the sacrifice of time for exhaustive fine-tuning or have poor performance in novel-class adaptation. We presume the major reason is that the valuable correlation feature among different categories is insufficiently exploited, hindering the generalization of knowledge from base to novel categories for object detection. In this paper, we propose few-shot object detection via Correlation-RPN and transformer encoder–decoder (CRTED), a novel training network to learn object-relevant features of inter-class correlation and intra-class compactness while suppressing object-agnostic features in the background with limited annotated samples. And we also introduce a four-way tuple-contrast training strategy to positively activate the training progress of our object detector. Experiments over two few-shot benchmarks (Pascal VOC, MS-COCO) demonstrate that our proposed CRTED without further fine-tuning can achieve comparable performance with current state-of-the-art fine-tuned works. The codes and pre-trained models will be released.

In the vast domain of computer vision, real-time object detection continues to occupy a paramount position. The intricate balancing act between accuracy and speed in object detectors presents a formidable hurdle for both academic scholars and industrial experts. Despite the remarkable proficiency displayed by recent transformer-based models in harnessing the attention mechanism, resulting in substantial performance gains over CNNs, their heavy computational demands often hinder their practical application in real-time detection environments. In this study, we have chosen YOLOv9 as our sturdy foundation and embarked on a journey of multi-module model refinement, ultimately culminating in the development of a novel, high-performing object detector dubbed MR-YOLOv9. To capitalize further on the representational prowess of feature images, we have opted for BiFPN as a formidable feature extractor, replacing the baseline's original FPN structure. This enhancement bolsters the efficiency and efficacy of feature fusion by strengthening the bidirectional flow of features, thereby elevating object detection performance. Moreover, we introduce a groundbreaking loss function, Inner-CIoU, which further sharpens the performance of our object detector when combined with the innovative Inner-IoU loss function. To validate our proposed methodologies, we have conducted extensive ablation studies on the MS-COCO 2017 detection datasets. Remarkably, our MR-YOLOv9 has been trained solely on the MS-COCO dataset, without leveraging any prior knowledge, yet it achieves an impressive 56.0% AP on the COCO 2017 test set, surpassing YOLOv9-E by a significant margin of 0.4% AP, all without incurring additional inference costs.

Deep learning has been one of the main trends in machine learning, and one of the most popular scientific research trends in recent years, which has played a revolutionary role in the development of computer vision. As a vital part of computer vision, many researchers have been attracted to participate in the research of image classification. With the emergence and development of deep neural networks, combined with the emerging various training and optimization methods, it has achieved great success in this field. In recent years, Transformer-based models has achieved the best experimental results in image classification and many other fields. This article will compactly introduce the development process of deep learning in image classification in the past few years from three aspects of training and optimization methods, deep neural networks and Transformers.

Deep learning is a key branch of machine learning that leads the frontier of scientific research and has a significant impact on the field of computer vision. With the development of deep learning, object detection, as a research hotspot in computer vision, has attracted a great number of researchers to conduct in-depth research. This paper introduces single-stage, two-stage and Transformer-based object detection algorithms and its backbone network structures, and provides a detailed comparison of the advantages and shortcomings of these algorithms. Finally, the practical applications of these algorithms in various fields and their performance evaluation on critical datasets are presented.

Robotic grasping is essential for intelligent robot operation. The performance of grasp detection relies on the measurement of the grasp rectangle's position and orientation. The rotation detector based on regression, however, faces the issue of boundary discontinuity as a result of the periodicity of the angle. This boundary discontinuity issue may make it difficult for the loss function to determine the precise grasp direction at the angle around the discontinuity. By comparing the differences between the conventional regression-based rotation, one-hot based rotation, and the CSL-based rotation in loss calculation, we introduce a circular smooth label (CSL) strategy to address the orientation boundary discontinuity problem and increase the error tolerance to adjacent angles. Based on the Yolov5 framework, we put our suggested approach into practice and ran tests on the Cornell grasp datasets. The experiment results indicate that the CSL-YOLOv5 outperforms the existing grasping methods on grasp rectangle predictions.

Considering the difficulty of modeling the individual head related transfer function (HRTF), a binaural moving sound source-oriented stereo audio synthesis approach is proposed, which refers to the interpolation method of HRTF. By means of spherical recording, the spatial audio of points from different directions can be obtained. In order to achieve more realistic effect, the spatial bilinear interpolation method is employed to calculate the weight of the relevant points which used to synthesize the moving sound source, and finally the Doppler Effect is simulated with the interpolation and extraction method in frequency domain. The experiment results show that, our method is capable of replacing HRTF to synthesize moving sound sources approximatively, and the generated performance of moving sound source is closer to the real recording.

This paper proposes a damage estimation algorithm from cues of image changes. We get the feature map of damage area through comparing the Haar feature matrix and the LBP feature matrix by two images before and after the change. We then take the offset comparison method for fusion comparison results of different migration. At last, we get accurate location of damage detection by Gaussian filter and image morphology processing. Experimental results show that the algorithm can accurately detect the image damage area effectively, and is not too sensitive for the changes of light and color temperature. Furthermore this method does not need to establish different damage detection and evaluation models for different targets, and it can adapt to a variety of conditions of damage detection.

Surveillance video is characterized by large amount of data and redundancy, which makes the suspicious face detection to be a problem. To solve the problem above we proposed suspicious face detection based on key frame. Surveillance video has the type of fixed background, so this paper used the frame difference method with low computational complexity and small computation to extract the key frame. We proposed a new method combined DPM with skin color detection to detect the suspicious face in the key frames. To solve the speed bottleneck of the traditional DPM, we proposed to use the fast HOG LUT feature extraction and the near optimal cost sensitive decision making improving the traditional method. Meanwhile we used YCrCb + otsu skin color segmentation. Since the otsu is easily affected by the illumination, we proposed an improved skin color detection. Experiment results show that the proposed algorithms are robust and accurate for real-time.