Region Proposal Generation Research Articles

Proposal-Free Fully Convolutional Network: Object Detection Based on a Box Map.

Region proposal-based detectors, such as Region-Convolutional Neural Networks (R-CNNs), Fast R-CNNs, Faster R-CNNs, and Region-Based Fully Convolutional Networks (R-FCNs), employ a two-stage process involving region proposal generation followed by classification. This approach is effective but computationally intensive and typically slower than proposal-free methods. Therefore, region proposal-free detectors are becoming popular to balance accuracy and speed. This paper proposes a proposal-free, fully convolutional network (PF-FCN) that outperforms other state-of-the-art, proposal-free methods. Unlike traditional region proposal-free methods, PF-FCN can generate a "box map" based on regression training techniques. This box map comprises a set of vectors, each designed to produce bounding boxes corresponding to the positions of objects in the input image. The channel and spatial contextualized sub-network are further designed to learn a "box map". In comparison to renowned proposal-free detectors such as CornerNet, CenterNet, and You Look Only Once (YOLO), PF-FCN utilizes a fully convolutional, single-pass method. By reducing the need for fully connected layers and filtering center points, the method considerably reduces the number of trained parameters and optimizes the scalability across varying input sizes. Evaluations of benchmark datasets suggest the effectiveness of PF-FCN: the proposed model achieved an mAP of 89.6% on PASCAL VOC 2012 and 71.7% on MS COCO, which are higher than those of the baseline Fully Convolutional One-Stage Detector (FCOS) and other classical proposal-free detectors. The results prove the significance of proposal-free detectors in both practical applications and future research.

SRFDet3D: Sparse Region Fusion based 3D Object Detection

Unlike the earlier 3D object detection approaches that formulate hand-crafted dense (in thousands) object proposals by leveraging anchors on dense feature maps, we formulate np (in hundreds) number of learnable sparse object proposals to predict 3D bounding box parameters. The sparse proposals in our approach are not only learnt during training but also are input-dependent, so they represent better object candidates during inference. Leveraging the sparse proposals, we fuse only the sparse regions of multi-modal features and we propose Sparse Region Fusion based 3D object Detection (SRFDet3D) network with mainly three components: an encoder for feature extraction, a region proposal generation module for sparse input-dependent proposals and a decoder for multi-modal feature fusion and iterative refinement of object proposals. Additionally for optimal training, we formulate our sparse detector with many-to-one label assignment based on Optimal Transport Algorithm (OTA). We conduct extensive experiments and analysis on publicly available large-scale autonomous driving datasets: nuScenes, KITTI, and Waymo. Our LiDAR-only SRFDet3D-L network achieves 63.1 mAP and outperforms the state-of-the-art networks on the nuScenes dataset, surpassing the dense detectors on KITTI and Waymo datasets. Our LiDAR-Camera model SRFDet3D achieves 64.7 mAP with improvements over existing fusion methods.

A deep learning-based and adaptive region proposal algorithm for semantic segmentation

This paper presents an adaptive and new region proposal algorithm for generating high-quality regions. The main aim of this algorithm is to investigate different features in the proposal generation process. This algorithm is based on bottom-up image segmentation and deep learning techniques. Each primary region is represented by features derived from a convolutional neural network (CNN). The adjacent and similar regions are merged based on a new proposed searching algorithm and a distance function in a hierarchical way. Additionally, various hand-crafted texture features are examined for representing each region. These texture features are not being utilized previously in region proposal generation. This method also applied both texture-based and deep learning-based features, contemporaneously. Furthermore, the proposed region proposal algorithm was evaluated on two significant challenging datasets, including VOC2012 and COCO2017. The resulting proposals show more high-quality regions in high overlaps in comparison to previous region proposal algorithms. More importantly, the results verified that the deep learning-based features together with handcrafted texture features are complement and this fusion can overcome the shortcoming of other approaches. Additionally, the proposed region proposal algorithm is employed in weakly supervised semantic segmentation. The new generated proposals are used to create some new labeled masks. These masks are very useful and effective in the training phase of deep learning. This approach is evaluated on VOC2012 and compared with previous region-based and free region-based methods. The results show the efficacy of the proposed method.

Microscopic hyperspectral imaging and an improved detection model based detection of Mycogone perniciosa chlamydospore in soil

The Mycogone perniciosa disease of Agaricus bisporus is highly contagious, with insignificant early symptoms and a long infestation period. Currently, there is a lack of a convenient and rapid means to detect the disease for early control. By using the transmission route of the disease, this paper focuses on M. perniciosa chlamydospore detection in contaminated soil. Microscopic hyperspectral images of M. perniciosa chlamydospore in contaminated soil were obtained to establish a detection model. Given a small target and a complex background, we proposed an improved spore detection model based on a faster regional convolutional neural network (Faster R-CNN). Furthermore, we combined the residual network Resnet50 and feature pyramid network (FPN) to extract thick spore target features at multiple scales. Meanwhile, we optimized the region proposal network (RPN) region proposal generation by adding two small scales to improve the performance of the detection model. The first three principal components (with 95% or more information) and RGB images were selected as model inputs, respectively, and the final average precision (AP) was 94.68% and 92.35%, respectively. This PC-based model also was compared to the VGG16 and Resnet50-based feature extraction networks of Faster R-CNN and Darknet53-based feature extraction network of the YOLOv3 model, and the AP was found to improve by 5.41%, 4.78%, and 6.34%, respectively. The results showed that micro-hyperspectral imaging combined with deep learning methods could accurately detect the chlamydospore in the soil, providing new methods and ideas for the early prevention and detection of M. perniciosa disease of A. bisporus.

Face mask detection and social distance monitoring system for COVID-19 pandemic.

Coronavirus triggers several respirational infections such as sneezing, coughing, and pneumonia, which transmit humans to humans through airborne droplets. According to the guidelines of the World Health Organization, the spread of COVID-19 can be mitigated by avoiding public interactions in proximity and following standard operating procedures (SOPs) including wearing a face mask and maintaining social distancing in schools, shopping malls, and crowded areas. However, enforcing the adaptation of these SOPs on a larger scale is still a challenging task. With the emergence of deep learning-based visual object detection networks, numerous methods have been proposed to perform face mask detection on public spots. However, these methods require a huge amount of data to ensure robustness in real-time applications. Also, to the best of our knowledge, there is no standard outdoor surveillance-based dataset available to ensure the efficacy of face mask detection and social distancing methods in public spots. To this end, we present a large-scale dataset comprising of 10,000 outdoor images categorized into a binary class labeling i.e., face mask, and non-face masked people to accelerate the development of automated face mask detection and social distance measurement on public spots. Alongside, we also present an end-to-end pipeline to perform real-time face mask detection and social distance measurement in an outdoor environment. Initially, existing state-of-the-art single and multi-stage object detection networks are fine-tuned on the proposed dataset to evaluate their performance in terms of accuracy and inference time. Based on better performance, YOLO-v3 architecture is further optimized by tuning its feature extraction and region proposal generation layers to improve the performance in real-time applications. Our results indicate that the presented pipeline performed better than the baseline version, showing an improvement of 5.3% in terms of accuracy.

Small Ship Detection Based on Hybrid Anchor Structure and Feature Super-Resolution

Small ships in remote sensing images have blurred details and are difficult to detect. Existing algorithms usually detect small ships based on predefined anchors with different sizes. However, limited by the number of different sizes, it is difficult for anchor-based methods to match small ships of different sizes and structures during training, as they can easily cause misdetections. In this paper, we propose a hybrid anchor structure to generate region proposals for small ships, so as to take full advantage of both anchor-based methods with high localization accuracy and anchor-free methods with fewer misdetections. To unify the output evaluation and obtain the best output, a label reassignment strategy is proposed, which reassigns the sample labels according to the harmonic intersection-over-union (IoU) before and after regression. In addition, an adaptive feature pyramid structure is proposed to enhance the features of important locations on the feature map, so that the features of small ship targets are more prominent and easier to identify. Moreover, feature super-resolution technology is introduced for the region of interest (RoI) features of small ships to generate super-resolution feature representations with a small computational cost, as well as generative adversarial training to improve the realism of super-resolution features. Based on the super-resolution feature, ship proposals are further classified and regressed by using super-resolution features to obtain more accurate detection results. Detailed ablation and comparison experiments demonstrate the effectiveness of the proposed method.

Surface Defect Detection with Modified Real-Time Detector YOLOv3

In this paper, a modified YOLOv3 net has been proposed for surface defect detection. Different from other pixel-level segmenting methods, YOLOv3 locates the regions of surface defects with bounding rectangles. Compared with conventional detectors, the operating efficiency of YOLOv3 is rather high without generating region proposals by sliding boxes. Although pixel-level details of defects are omitted in the process, the primary information of the location of detects and class labels are extracted by YOLOv3 with high accuracy. This information is sufficient for surface defect inspection, and computational efficiency has been improved, simultaneously. To further light the structure of YOLOv3, loss function optimization and pruning strategy have been adopted in the original YOLOv3. The pruning ratio is determined by the tradeoff between detecting accuracy and computational efficiency. In our experiments, we compared the performance of modified YOLOv3 with several state-of-the-art methods, and modified YOLOv3 achieves the best performance on six types of surface defects in DAGM 2007 dataset.

Vehicle Detection for Unmanned Systems Based on Multimodal Feature Fusion

This paper proposes a 3D vehicle-detection algorithm based on multimodal feature fusion to address the problem of low vehicle-detection accuracy in unmanned system environment awareness. The algorithm matches the coordinate relationships between the two sensors and reduces sampling errors by combining the millimeter-wave radar and camera calibration. Statistical filtering is used to remove redundant points from the millimeter-wave radar data to reduce outlier interference; a multimodal feature fusion module is constructed to fuse the point cloud and image information using pixel-by-pixel averaging. Moreover, feature pyramids are added to extract fused high-level feature information, which is used to improve detection accuracy in complex road scenarios. A feature fusion region proposal structure was established to generate region proposals based on the high-level feature information. The vehicle detection results were obtained by matching the detection frames in their vertices after removal of the redundant detection frames using non-maximum suppression. Experimental results from the KITTI dataset show that the proposed method improved the efficiency and accuracy of vehicle detection with the corresponding average of 0.14 s and 84.71%.

Generating impact maps from bomb craters automatically detected in aerial wartime images using marked point processes

Even more than 75 years after the Second World War, numerous unexploded bombs (duds) linger in the ground and pose a considerable hazard to society. The areas containing these duds are documented in so-called impact maps, which are based on locations of exploded bombs; these locations can be found in aerial images taken shortly after bombing. To generate impact maps, in this paper we present a novel approach based on marked point processes (MPPs) for the automatic detection of bomb craters in such images, some of which are overlapping. The object model for the craters is represented by circles and is embedded in the MPP-framework. By means of stochastic sampling, the most likely configuration of objects within the scene is determined. Each configuration is evaluated using an energy function that describes the consistency with a predefined object model. High gradient magnitudes along the object borders and homogeneous grey values inside the objects are favoured, while overlaps between objects are penalized. Reversible Jump Markov Chain Monte Carlo sampling, in combination with simulated annealing, provides the global optimum of the energy function. Our procedure allows the combination of individual detection results covering the same location. Afterwards, a probability map for duds is generated from the detections via kernel density estimation and areas around the detections are classified as contaminated, resulting in an impact map. Our results, based on 74 aerial wartime images taken over different areas in Central Europe, show the potential of the method; among other findings, a clear improvement is achieved by using redundant image information. We also compared the MPP method for bomb crater detection with a state-of-of-the-art convolutional neural network (CNN) for generating region proposals; it turned out that the CNN outperforms the MPPs if a sufficient amount of representative training data is available and a threshold for a region to be considered as crater is properly tuned prior to running the experiments. If this is not the case, the MPP approach achieves better results.

Scene captioning with deep fusion of images and point clouds

• Meaningful scene captions can be generated using the fusion of point clouds and images. • Features of region proposals from point clouds helps better understanding visual scenes. • Region proposal fusion and feature fusion can integrate multi-modal vision information. • Multi-modal vision fusion leads to better scene captions than single-modal vision data. Recently, the fusion of images and point clouds has received appreciable attentions in various fields, for example, autonomous driving , whose advantage over single-modal vision has been verified. However, it has not been extensively exploited in the scene captioning task. In this paper, a novel scene captioning framework with deep fusion of images and point clouds based on region correlation and attention is proposed to improve performances of captioning models. In our model, a symmetrical processing pipeline is designed for point clouds and images. First, 3D and 2D region features are generated respectively through region proposal generation, proposal fusion, and region pooling modules. Then, a feature fusion module is designed to integrate features according to the region correlation rule and the attention mechanism, which increases the interpretability of the fusion process and results in a sequence of fused visual features. Finally, the fused features are transformed into captions by an attention-based caption generation module. Comprehensive experiments indicate that the performance of our model reaches the state of the art.

SFGNet detecting objects via spatial fine-grained feature and enhanced RPN with spatial context

Object detection, which is one of the most fundamental visual recognition tasks, has been a hotspot in computer vision. CNN (Convolutional Neural Networks) have been widely employed for building detector. Due to the success of RPN (Region Proposal Network), the two-stage detectors get both classification accuracy and precise regression bounding boxes. However, they still struggle in small-size object detection. In this paper, we present a deep network, namely Spatial Fine-Grained Network (SFGN). The SFGN that exploits Spatial Fine-Grained Features (SFGF) concatenates the higher resolution features, which is fine-grained with the low resolution features and high-level semantic by stacking spatial features for fine-grained features. An enhanced region proposal generator is proposed to get the objectless for small object to obtain a small set of proposal. The contextual information surrounding the region of interest is embedded using local spatial information for increasing the useful information and discriminating the background. For improving the detection performance, we use a simple yet surprisingly effective online hard example mining (OHEM) algorithm for training region proposal generator. It embeds an efficiently implemented soft non-maximum suppression (soft-NMS) for replacing with tradition NMS to obtain consistent improvements without increasing the computational complexity in inference. On PASCAL VOC 2007 and PASCAL VOC 2012 datasets, our SFGN improves baseline model from 81.2% mAP to 80.6% mAP. On MS COCO dataset, SFGN also performs better than baseline model. As intuition suggests, our detection results provide strong evidence that our SFGN improves detection accuracy, especially in small object test.

Detection with fast feature pyramids and lightweight convolutional neural network: a practical aircraft detector for optical remote images

Aircraft detection in optical remote sensing images has important practical value and attracts great attention in recent years. Due to some clustered backgrounds and various deformations like variability in viewpoints, scales, and directions, it is still a big challenge. In recent years, deep learning, especially convolutional neural networks (CNNs), has achieved remarkable success not only in natural image object detection tasks but also in remote sensing image object detection tasks. However, due to the limitations of the actual application deployment environment only limited processing units and storage space can be used, and most deep learning-based aircraft detection algorithms cannot work well. To this end, a unique aircraft detection framework in optical remote sensing images is proposed, which is based on fast feature pyramids and a specialized lightweight CNN. The algorithm can be summarized into two steps. In the first step, a highly reliable region proposal generation algorithm is designed to predict region proposals. In this phase, a prior square sliding window method and an efficient linear search algorithm are proposed. In the next step, a lightweight CNN is designed to achieve fast and accurate aircraft classification. The proposed model only requires about 237 KB of parameter storage and can infer in a poor central processing unit platform. Comprehensive experiments on three publicly optical remote sensing aircraft datasets demonstrate the superiority and effectiveness of the proposed method. It is also shown that the proposed method is light and fast in comparison with some state-of-the-art methods.

Enhancing representation learning by exploiting effective receptive fields for object detection

Most of state-of-the-art object detectors depend on multiple anchors/reference boxes in representation learning. However, such anchor-based representation does not completely match with the visual information perceived by the sliding windows, thus degrading the overall performance of object detection. In this paper, we present an effective receptive field (eRF)-dependent region proposal network (eRPN) for proposal generation, which enhances the anchor-based representation via eRFs. Specifically, we define an eRF for each sliding window on the feature map and only encode objects within the eRF for unbiasedly representation learning. The size of eRF depends on its backbone network. An eRF-based matching rule is devised and combined with the commonly used IoU rule for pertinent sample selection. We also design an eRF filter module, which can be appended to RPN for eliminating redundant low-quality region proposals in inference time. eRPN enhances representation learning from two perspectives: input information and sample balance, to make generating region proposals more robust. We evaluate eRPN by combining with two commonly used detection heads: Faster RCNN and Faster RCNN w FPN(Faster-FPN). Experimental results on PASCAL VOC dataset and MS COCO dataset benchmarks demonstrate the effectiveness of the proposed method in learning representation for object detection.

Region proposal patch-image model for infrared small target detection

ABSTRACT The infrared patch-image (IPI) model has been proved to be very successful in infrared small target detection. However, the IPI model suffers from two major issues: abundant of residuals in the recovered target image and low detection speed caused by the large-size low-rank decomposition matrix. To mitigate these challenges, a novel infrared small target detection method called the region proposal patch-image (RPPI) model is proposed. The method contains two stages, where the first one is to generate region proposals (RPs) based on the contrast mechanism and the second one is to recover targets based on matrix decomposition. In the first stage, RPs containing interest targets are extracted in order to suppress background clutters and reduce the size of the decomposed matrix. In RP extraction, the contrast mechanism instead of the original convolutional neural network (CNN) is adapted due to lack of texture features for infrared small target and high computational complexity with CNN. In the second stage, considering the weakness of the correlation between RPs, a weighted nuclear norm (WNN) is utilized for avoiding relaxation of the sparsity too much, which can preserve targets and suppress the background simultaneously. Finally, accelerated proximal gradient (APG)-combined WNN is applied to solve this model. Experimental results on eight image groups with various kinds of targets and backgrounds demonstrate that the proposed method can effectively detect the target with faster speed, higher detection rate, and lower false alarm rate compared with other state-of-the-art methods.

Anti-Motion Remote Measurement of Heart Rate Based on Region Proposal Generation and Multi-Scale ROI Fusion

Remote detection of human heart rate (HR) in a non-contact manner has become an active research topic in recent years. Most existing works focus on signal separation of raw traces obtained from the facial region of interest (ROI), whereas the ROI itself has a crucial impact on the measurement accuracy. The ROI tracking issues such as not fitting the subject’s face, pixel shift when there is head motion, and fluctuation when keeping still cannot be effectively dealt with by existing remote photoplethysmography (rPPG) algorithms, which seriously affects the pulse measurement accuracy. This article addresses these issues by proposing a novel pulse signal extraction framework based on generating region proposals and multi-scale ROIs. The framework consists of two signal selection modules and one signal fusion module. First, the multiple aspect ratio sampling operation is proposed to generate bounding boxes with various aspect ratios to remove background pixels as well as to fit different subjects’ faces. Second, the bounding box perturbation strategy is proposed to generate bounding boxes around the selected position to reduce the tracking error caused by head translation and rotation. Finally, multi-scale signal fusion (MSSF) is proposed for fusion signals from multi-scale ROI. Extensive experiments are conducted on two publicly available datasets and one self-collected dataset. The results show that the proposed framework achieves significant performance improvement in a signal-to-noise ratio (SNR), mean absolute error (MAE), and root mean squared error (RMSE). On pulse rate detection dataset (PURE), Université Bourgogne Franche-Comté (UBFC)-rPPG, and Self-rPPG datasets, the SNR has a maximum increase of 2.848 dB [compared with chrominance-based rPPG (CHROM)], 1.690 dB [compared with plane-orthogonal-to-skin (POS)], and 1.806 dB (compared with CHROM), respectively. MAE has a maximum decrease of 1.947 bpm (compared with CHROM), 1.610 bpm [compared with pulse blood volume (PBV)], and 1.577 bpm (compared with CHROM), respectively. Moreover, the proposed framework is independent of existing signal extraction algorithms. They can be easily embedded in the proposed framework.

Fabric defect detection with an attention mechanism based on hard sample training

In view of the various types of fabric defects, and the problems of confusion, density unevenness and small target defects, which are difficult to detect, this paper builds a deep learning defect detection network incorporating an attention mechanism. The data augmentation strategy is used to enrich the number of samples of each defective type, and the enriched samples were extracted by the feature extraction network integrated with the attention mechanism, which can improve the feature extraction ability of confusable defect types and small defect types. Region proposal generation generates a proposal box for extracted features, and adds an online hard example mining strategy to re-learn hard examples to accelerate network convergence. Region feature aggregation maps the proposal box to the feature map to obtain the region of interest. Finally, the defect features are classified and the bounding boxes are regressed. The results show that this algorithm can effectively detect 39 categories of fabric defects with a detection speed of 0.085 s and a detection accuracy of 0.9338.

A Multi-Scale Spatial Attention Region Proposal Network for High-Resolution Optical Remote Sensing Imagery

Deep convolutional neural networks (DCNNs) are driving progress in object detection of high-resolution remote sensing images. Region proposal generation, as one of the key steps in object detection, has also become the focus of research. High-resolution remote sensing images usually contain various sizes of objects and complex background, small objects are easy to miss or be mis-identified in object detection. If the recall rate of region proposal of small objects and multi-scale objects can be improved, it will bring an improvement on the performance of the accuracy in object detection. Spatial attention is the ability to focus on local features in images and can improve the learning efficiency of DCNNs. This study proposes a multi-scale spatial attention region proposal network (MSA-RPN) for high-resolution optical remote sensing imagery. The MSA-RPN is an end-to-end deep learning network with a backbone network of ResNet. It deploys three novel modules to fulfill its task. First, the Scale-specific Feature Gate (SFG) focuses on features of objects by processing multi-scale features extracted from the backbone network. Second, the spatial attention-guided model (SAGM) obtains spatial information of objects from the multi-scale attention maps. Third, the Selective Strong Attention Maps Model (SSAMM) adaptively selects sliding windows according to the loss values from the system’s feedback, and sends the windowed samples to the spatial attention decoder. Finally, the candidate regions and their corresponding confidences can be obtained. We evaluate the proposed network in a public dataset LEVIR and compare with several state-of-the-art methods. The proposed MSA-RPN yields a higher recall rate of region proposal generation, especially for small targets in remote sensing images.

A self-attention based faster R-CNN for polyp detection from colonoscopy images

At present, the incidence rate of colorectal cancer (CRC) is increasing year by year. It has always affected people's physical and mental health and quality of life. How to improve the detection ability of polyp plays a key role in colonoscopy. In order to solve these problems, in this paper, we first enhance the contrast of the input image by well distinguishing the foreground from the background in order to improve the saliency of the polyp regions. Then, we feed the enhanced image into an improved Faster R-CNN architecture comprised three processing modules for feature extraction, region proposal generation, and polyp detection, respectively. In order to further improve the quality, as well as the feature abstraction capability of the feature maps produced by the feature extraction network, we append an attention module to attend to the useful feature channels and weaken the contributions of the helpless feature channels. The experimental results demonstrate that the accuracy of the proposed polyp detection network is greatly improved compared with the existing algorithms, and the network not only can accurately identify polyps of varying sizes and conditions in single polyp images, but also can achieve excellent performance in handling multiple polyp images. This paper will be greatly helpful to alleviate the missed diagnosis of clinicians in the process of endoscopic examination and disease treatment, as well as providing effective assistance for the early diagnosis, treatment and prevention of the CRC, which is also of great significance to the clinical work of physicians.

An Enhanced Insect Pest Counter Based on Saliency Map and Improved Non-Maximum Suppression.

Simple SummarySimple Summary: Accurately counting the number of insect pests from digital images captured on yellow sticky traps remains a challenge in the field of insect pest monitoring. This paper develops a new approach to counting the number of insect pests using a saliency map and improved non-maximum suppression. Specifically, a saliency map is exploited to construct a region proposal generator, and a convolutional neural network (CNN) model is used to classify each region proposal as a specific insect pest class, resulting in detection bounding boxes. An improved non-maximum suppression is further developed to sophisticatedly handle the redundant detection bounding boxes, and the insect pest number is thus obtained through counting the handled detection bounding boxes, each of which covers one insect pest. As this insect pest counter may miscount insect pests that are close to each other, the widely used Faster R-CNN is further integrated with the mentioned insect pest counter to construct a dual-path network. Extensive experimental simulations show that the two proposed insect pest counters achieve significant improvements in terms of F1 score against state-of-the-art object detectors as well as insect pest detection methods.Accurately counting the number of insect pests from digital images captured on yellow sticky traps remains a challenge in the field of insect pest monitoring. In this study, we develop a new approach to counting the number of insect pests using a saliency map and improved non-maximum suppression. Specifically, as the background of a yellow sticky trap is simple and the insect pest object is small, we exploit a saliency map to construct a region proposal generator including saliency map building, activation region formation, background–foreground classifier, and tune-up boxes involved in region proposal generation. For each region proposal, a convolutional neural network (CNN) model is used to classify it as a specific insect pest class, resulting in detection bounding boxes. By considering the relationship between detection bounding boxes, we thus develop an improved non-maximum suppression to sophisticatedly handle the redundant detection bounding boxes and obtain the insect pest number through counting the handled detection bounding boxes, each of which covers one insect pest. As this insect pest counter may miscount insect pests that are close to each other, we further integrate the widely used Faster R-CNN with the mentioned insect pest counter to construct a dual-path network. Extensive experimental simulations show that the two proposed insect pest counters achieve significant improvement in terms of score against the state-of-the-art object detectors as well as insect pest detection methods.

Moving Object Detection for Airport Scene Using Patterns of Motion and Appearance

This paper presents a novel method for localization and recognition of moving objects in a real airport surface scene. Different from the traditional applications, moving object detection (MOD) in the airport surface is more challenging because the background is an open outdoor environment, which means that the target objects are usually low in resolution and the MOD task is vulnerable to many undesired changes, such as cloud movement and illumination variations. To address these issues, this paper proposes a unified and effective deep-learning-based MOD architecture, which combines both appearance and motion cues. Specifically, a novel moving region proposal generation module is first designed, which can effectively locate the regions of moving object based on the motion information. Meanwhile, a novel cascade multilayer feature fusion module with transposed convolution is applied to produce both enriched-semantics and fine-resolution convolutional feature maps for category recognition. Finally, a large-scale dataset acquired by the daily surveillance videos of a real airport surface is manually constructed. Results show that the proposed methods outperform state-of-the-art solutions in extracting moving objects from airport surface scenes.