Underwater object detection using side-scan sonar (SSS) remains a significant challenge in marine exploration, especially for small objects. Conventional methods face various obstacles, such as difficulties feature extraction and the considerable impact of noise on accuracy. To address these issues, this study proposes an improved YOLOv11 network named YOLOv11-SDC. Specifically, new Sparse Feature (SF) module is proposed, replacing Spatial Pyramid Pooling Fast (SPPF) from original...

Underwater object detection is an important task in marine exploration. The existing autonomous underwater vehicle (AUV) designs typically lack integrated module and are constrained by communication limitations environments. This results a situation where AUV, when tasked with missions, require real-time transmission of sensing data to shore-based stations but unable do so. Consequently, the divided into two discontinuous phases: AUV acquisition detection, leading limited autonomy...

Palm vein recognition is a high-security biometric. Outside the NIR capture box and contactless palm are more popular but challenging. The users feel comfortable outside face optical blurring brought by visible light. Contactless gestures solve hygienic problem image rotation, position translation, scale transformation which makes classification difficult especially in large-scale databases. To address these problems, we develop wavelet denoising ResNet, consists of two models: (WD) model...

Underwater object recognition in sonar images, such as mine detection and wreckage of a submerged airplane, is very challenging task. The main difficulties include but are not limited to rotation, confusion from false targets complex backgrounds, extensibility ability on diverse types objects. In this paper, we propose an underwater method using transformable template matching approach based prior knowledge. Specifically, first extract features construct video sequences the analysis acoustic...

Diver target automatic detection is indispensable for underwater defense systems, such as the unmanned harbor surveillance system. It a very challenging task due to various poses and intensity features of diver target. In addition, background noise in sonar images complex, which also makes more difficult. this paper, we propose method based on saliency images. On basis studying characteristics images, first decompose original image perform median filtering it, can significantly improve...

Recently, side scan sonar (SSS) is increasingly applied to underwater search, which can display the microgeomorphic morphology and distribution, obtain a continuous two-dimensional submarine acoustic map with certain width. Automatic object detection methods help lot in case of long searches, where operators may feel exhausted therefore miss possible object. This paper proposes an method based on YOLO-v3 network. We first establish real image data-set, includes 7000 images four types...

Deep learning is recognized to be capable of discovering deep features for representation and pattern recognition without requiring elegant feature engineering techniques by taking advantages human ingenuity prior knowledge. Thus it has triggered enormous research activities in machine recognition. One the most important challenges figure out relations between a depth neural networks (deep nets short) reflect necessity depth. Our purpose quantify this feature-depth correspondence extraction...

Image clustering is a complex procedure, which significantly affected by the choice of image representation. Most existing methods treat representation learning and separately, usually bring two problems. On one hand, representations are difficult to select learned not suitable for clustering. other they inevitably involve some step, may error hurt results. To tackle these problems, we present new method that efficiently builds an precisely discovers cluster assignments. For this purpose,...

In this paper, we present an automatic system of mine like object detection and recognition for sonar videos. This is implemented with two main methods. One the segmentation intrackability, another based on improved BOW algorithm Support Vector Machine (SVM). Intrackability defined by concept entropy, can reflect difficulty uncertainty in tracking certain elements time axis. Therefore, our method effectively eliminate complex noise image to guarantee more accurate detection. object, use SVM...

Image clustering is a fundamental problem in computer vision domains. In this survey, we provide comprehensive overview of image clustering. Specifically, first discuss the applications across various Then, summarize common algorithms and propose classification The existing methods are classified from four aspects: autoencoder based methods, subspace clustering, graph convolution network (GCN) some other methods. We introduce main research contents problems also recent experimental results....

There exist various methods for transferring knowledge between neural networks, such as parameter transfer, feature sharing, and distillation. However, these are typically applied when networks of equal size or from larger to smaller ones. Currently, there is a lack shallower deeper ones, which crucial in real-world scenarios system upgrades where network increases better performance. End-to-end training the commonly used method training. this strategy, cannot inherit existing network. As...

This paper proposes a sparse video representation with deformable spatiotemporal template feature, named as active trace template. An is the motion track of an spatial which moves in certain velocity. To accommodate geometric variations feature and temporal track, atomic each frame capable to slightly shift its location other attributes within ranges best represent salient trackable structure. The quality quantified by score. It computed through new proposed hierarchical architecture sum-max...

Image clustering is a complex procedure that significantly affected by the choice of image representation. Generally speaking, representations are generated using handcraft features or trained neural networks. When dealing with high dimension data, these two representation methods cause problems: i) ability manually designed limited; ii) non-representative and meaningless feature deep network may hurt performance. To overcome problems, we propose new method which efficiently builds an...

Deep learning is recognized to be capable of discovering deep features for representation and pattern recognition without requiring elegant feature engineering techniques by taking advantage human ingenuity prior knowledge. Thus it has triggered enormous research activities in machine recognition. One the most important challenge figure out relations between a depth neural networks (deep nets short) reflect necessity depth. Our purpose quantify this feature-depth correspondence extraction...

Abstract Object detection usually includes two parts: objection classification and location. At present, the popular object detectors use heads: one head is used to predict score, other bounding box (bbox), respectively. In this paper, we first stack after feature extract convolutional neural networks of bbox regression head. Then, establish by using a feature. The very useful when uses soft Intersection over Union (IoU) labels. experiment parts, only PASCAL VOC 2007 datasets, Centerness...