Discriminative Feature Representation Research Articles

SnipeDet: Attention-guided pyramidal prediction kernels for generic object detection

• APPK module is proposed to tackle misalignments of different scales of objects. • IoU-adaptive loss function helps networks to deal with the hard negative samples. • SORR module is devised to improve the detection efficiency. • Interleaved subsampling method can enhance feature representations. Using single-scale prediction kernels or Region of Interest (RoI) pooling in the prediction modules of modern object detectors is not very successful in matching different scales of objects. State-of-the-art detectors with the feature pyramid structure built on different resolutions of feature maps can help alleviate this problem. Although with this structure, single-scale prediction kernels or RoI pooling still struggles to detect small objects, and simultaneously, the former continues to encounter the misalignment problem on very large objects. In this paper, we propose the attention-guided pyramidal prediction kernels module with a customized IoU-adaptive loss function to deal with the misalignment problem between the prediction module and different scales of objects. To mitigate the effect of heavy detection head, we also introduce the salient object regions recognition module to identify these regions that have strong object cues. Additionally, interleaved subsampling, as the proposed feature enhancement approach, is used to generate highly discriminative feature representations. We refer to the detection framework constituted by these proposed methods as SnipeDet. Results show that SnipeDet achieves 41.1 AP at the speed of 15.4 FPS on the MS COCO test-dev set with 512 × 512 input images, which outperforms state-of-the-art one-stage detectors and has a better trade-off between speed and accuracy.

Perceiving informative key-points: A self-attention approach for person search

Though person re-identification has witnessed significant progress, person search, as a more practical task considering the unavailability of annotations of pedestrian bounding boxes, has progressed much slower because of less discriminative feature representation. To this end, we propose a novel self-attention based person search approach by perceiving informative implicit key-points with weak supervision. Firstly, we design the Self-attention Slice Part (SSP) module, to implicitly localize informative key-points by only taking a pre-defined number of points as supervision. Concretely, this module utilizes both channel-attention and spatial attention with weak supervision on partitioned pedestrian slices to get the most discriminative key-points. After that, we strengthen the self-attention weight for these cardinal key-points, and then, more robust feature representations for conducting person search can be obtained from these self-mined key-points. Meanwhile, the SSP also provides semantic alignment with the horizontally partitioned slices. Besides, to focus more on reducing the inner-class margin rather than enlarging inter-class distance, the Random Label Smooth (RLS) loss is defined for more robust classification. The RLS loss not only provides a larger margin hyperplane but also enhances the training efficiency. Therefore, all in all: (1) We propose an end-to-end person search framework by fully exploiting current object detection and person re-identification techniques jointly with weak supervision. (2) Our proposed weakly-supervised self-attention module is generic and can be plugged into any related tasks to improve the performance. (3) We conduct extensive experiments on popular benchmarks, including the dataset of CUHK-SYSU and PRW, and our approach outperforms most current state-of-the-art methods according to mAP and top-1 evaluation metrics.

Transfer Learning for Alzheimer's Disease through Neuroimaging Biomarkers: A Systematic Review.

Alzheimer’s disease (AD) is a remarkable challenge for healthcare in the 21st century. Since 2017, deep learning models with transfer learning approaches have been gaining recognition in AD detection, and progression prediction by using neuroimaging biomarkers. This paper presents a systematic review of the current state of early AD detection by using deep learning models with transfer learning and neuroimaging biomarkers. Five databases were used and the results before screening report 215 studies published between 2010 and 2020. After screening, 13 studies met the inclusion criteria. We noted that the maximum accuracy achieved to date for AD classification is 98.20% by using the combination of 3D convolutional networks and local transfer learning, and that for the prognostic prediction of AD is 87.78% by using pre-trained 3D convolutional network-based architectures. The results show that transfer learning helps researchers in developing a more accurate system for the early diagnosis of AD. However, there is a need to consider some points in future research, such as improving the accuracy of the prognostic prediction of AD, exploring additional biomarkers such as tau-PET and amyloid-PET to understand highly discriminative feature representation to separate similar brain patterns, managing the size of the datasets due to the limited availability.

A discriminant kernel entropy-based framework for feature representation learning

The intelligent multimedia processing community has developed increasing interest in kernel entropy component analysis (KECA) due to its abilities in effective data transformation and dimensionality reduction. However, since only the unsupervised structural information of Renyi entropy from the given data set is utilized, KECA alone is incapable of generating high quality discriminant features. Aiming to develop a new and generic approach for feature representation learning, this paper proposes a discriminant kernel entropy-based framework, which integrates KECA and a complete discriminant strategy (consisting of regular and irregular discriminant information), to explore discriminant feature representations from the given data set. The framework is realized and further optimized to generate a more powerful discriminant descriptor for feature representation learning, leading to improved performance. Since the joint utilization of kernel entropy estimation and the complete discriminant strategy is able to reveal the distribution and semantic information of the given data, the proposed framework opens up a new front for discriminant feature representation learning via information theoretic learning (ITL). To demonstrate the generic nature and effectiveness of the proposed framework, experiments are conducted on two different data sources; the visual data source (e.g., University of California Irvine (UCI) database, Olivetti Research Lab (ORL) database, Caltech 256 database) and the audio data source (Ryerson Multimedia Lab (RML) audio emotion database). The results show this framework yields superior performance over other methods on the data sets evaluated for feature representation learning.

Semi-supervised learning for person re-identification based on style-transfer-generated data by CycleGANs

Person re-identification (re-ID) is an exceedingly significant branch in the field of computer vision, especially for video surveillance. It is still a challenge to obtain more labeled training data and use them reasonably for more precise matching, though the person re-ID performance has been improved significantly. In order to solve this challenge, this study proposes a semi-supervised learning algorithm for data augmentation, the style-transfer-generated data as an extra class (STGDEC), which is aided by the Cycle-Consistent Adversarial Networks (CycleGANs) in generating extra unlabeled training data. Specifically, the algorithm firstly trains the CycleGANs and Deep Convolutional Generative Adversarial Networks so as to generate large amounts of unlabeled data. Secondly, we propose an adaptive receptive field module to expand the size of receptive fields and select the appropriate receptive field features dynamically in order to learn more contextual information and discriminative feature representation and embed the module in the backbone network easily. Thirdly, we use the combination of label smoothing regularization for outliers and an extra class loss to regularize the generated data and encourage the network not to be too confident to the ground-truth. Finally, this paper proposes three training strategies for the combination of standard dataset and generated samples. Comprehensive experiments based on the STGDEC are conducted, and these results show that the proposed algorithm gains a significant improvement over the baseline, the Basel. + LSRO and state-of-the-art approaches of person re-ID in many cases.

Image Synthesis Based On Feature Description

Abstract: Generating realistic images from text is innovative and interesting, but modern-day machine learning models are still far from this goal. With research and development in the field of natural language processing, neural network architectures have been developed to learn discriminative text feature representations. Meanwhile, in the field of machine learning, generative adversarial networks (GANs) have begun to generate extremely accurate images of especially in categories, such as faces, album covers, and room interiors. In this work, the main goal is to develop a neural network to bridge these advances in text and image modelling, by essentially translating characters to pixels the project will demonstrate the capability of generative models by taking detailed text descriptions and generate plausible images. Keywords: Deep Learning, Computer Vision, NLP, Generative Adversarial Networks

Blind image quality assessment with channel attention based deep residual network and extended LargeVis dimensionality reduction

Image Quality Assessment (IQA) is one of the fundamental problems in the fields of image processing, image/video coding and transmission, and so on. In this paper, a Blind Image Quality Assessment (BIQA) approach with channel attention based deep Residual Network (ResNet)and extended LargeVis dimensionality reduction is proposed. Firstly, ResNet50 with channel attention mechanism is used as the backbone network to extract the deep features from the image. In order to reduce the dimensionality of the deep features, LargeVis, which is originally designed for the visualization of large scale high-dimensional data, is extended by using Support Vector Regression (SVR) to perform on a single feature vector data. The extended LargeVis can remove the redundant information of the deep features so as to obtain a low-dimensional and discriminative feature representation. Finally, the quality prediction model is established by using SVR as the fitting method. The low-dimensional feature representation and quality score of the image form the pair-wise data samples to train the fitting model. Experimental results on authentic distortions datasets and synthetic distortions datasets show that our proposed method can achieve superior performance compared with the state-of-the-art methods.

A novel radar target recognition method for open and imbalanced high-resolution range profile

High-resolution range profile (HRRP) data often has imbalanced and open-ended distribution in realistic radar target recognition task, requiring the recognition system classify targets among majority and minority classes and detect unseen targets accurately. In this paper, we propose a novel algorithm for open and imbalanced HRPP recognition tasks, by learning from realistic data distribution and optimizing the accuracy over the majority, minority and open classes. This method maps the HRRP data to the latent feature space, enhances the feature by sharing knowledge from majority to minority class, and provides a scalar indicating the familiarity to known classes. Dual-attention is developed to provide strong discriminative feature representation. An angular penalty is employed in the loss function to optimize the intra-class similarity and inter-class variability. Experiments on measured data prove that the proposed algorithm outperforms other existing methods under both close and open settings with accuracy increased significantly, and more discriminative representation. This study provides a promising and effective approach for open and imbalanced HRRP target recognition.

Split-Attention Networks with Self-Calibrated Convolution for Moon Impact Crater Detection from Multi-Source Data

Impact craters are the most prominent features on the surface of the Moon, Mars, and Mercury. They play an essential role in constructing lunar bases, the dating of Mars and Mercury, and the surface exploration of other celestial bodies. The traditional crater detection algorithms (CDA) are mainly based on manual interpretation which is combined with classical image processing techniques. The traditional CDAs are, however, inefficient for detecting smaller or overlapped impact craters. In this paper, we propose a Split-Attention Networks with Self-Calibrated Convolution (SCNeSt) architecture, in which the channel-wise attention with multi-path representation and self-calibrated convolutions can generate more prosperous and more discriminative feature representations. The algorithm first extracts the crater feature model under the well-known target detection R-FCN network framework. The trained models are then applied to detecting the impact craters on Mercury and Mars using the transfer learning method. In the lunar impact crater detection experiment, we managed to extract a total of 157,389 impact craters with diameters between 0.6 and 860 km. Our proposed model outperforms the ResNet, ResNeXt, ScNet, and ResNeSt models in terms of recall rate and accuracy is more efficient than that other residual network models. Without training for Mars and Mercury remote sensing data, our model can also identify craters of different scales and demonstrates outstanding robustness and transferability.

Deep subclass reconstruction network for fault diagnosis of rotating machinery under various operating conditions

In the real-world industrial application, all measured vibration signals are usually suffered from substantial spurious factors and characterized as large intra-class and intra-subclass variations due to changes in the operating conditions (i.e., working loads, shaft speeds), which seriously deteriorates most of the machine learning methods’ ability to learn the discriminative feature representations. In this work, we propose a novel subclass reconstruction network (SCRN) to learn discriminative feature representations from raw vibration signals under different working conditions by suppressing the intra-class and intra-subclass variations in the feature space. Specifically, a novel and simple average strategy is developed to represent the cluster centroid of each subclass information as an effective supervised representation which is effectively embedded into the SCRN model. Furthermore, a new cost function of SCRN is formulated by jointly minimizing the basic and subclass-level reconstruction errors. To better exploit the discriminative information and improve classification performance, we further develop a deep subclass reconstruction network (DSCRN) model by stacking multiple SCRN models together through non-linear transformations to learn better deep feature representations. Extensive comparative evaluations on three benchmarks CWRU, 2009 PHM, and MFPT demonstrate that the proposed methods achieve consistently better diagnosis performance than existing state-of-the-art methods. Especially, when only 1% training samples from CWRU are used, the proposed DSCRN has significant improvement with classification performance of 94.55% over compared method with a gain of at least 9%.

Deep cross-view co-regularized representation learning for glioma subtype identification

The new subtypes of diffuse gliomas are recognized by the World Health Organization (WHO) on the basis of genotypes, e.g., isocitrate dehydrogenase and chromosome arms 1p/19q, in addition to the histologic phenotype. Glioma subtype identification can provide valid guidances for both risk-benefit assessment and clinical decision. The feature representations of gliomas in magnetic resonance imaging (MRI) have been prevalent for revealing underlying subtype status. However, since gliomas are highly heterogeneous tumors with quite variable imaging phenotypes, learning discriminative feature representations in MRI for gliomas remains challenging. In this paper, we propose a deep cross-view co-regularized representation learning framework for glioma subtype identification, in which view representation learning and multiple constraints are integrated into a unified paradigm. Specifically, we first learn latent view-specific representations based on cross-view images generated from MRI via a bi-directional mapping connecting original imaging space and latent space, and view-correlated regularizer and output-consistent regularizer in the latent space are employed to explore view correlation and derive view consistency, respectively. We further learn view-sharable representations which can explore complementary information of multiple views by projecting the view-specific representations into a holistically shared space and enhancing via adversary learning strategy. Finally, the view-specific and view-sharable representations are incorporated for identifying glioma subtype. Experimental results on multi-site datasets demonstrate the proposed method outperforms several state-of-the-art methods in detection of glioma subtype status.

MSGSE-Net: Multi-scale guided squeeze-and-excitation network for subcortical brain structure segmentation

Convolutional neural networks (CNNs) have been achieving remarkable results in medical image segmentation. However, for accurate segmentation of subcortical brain structure in MR images, it is still a challenge due to the ambiguous boundaries, the complex structures, and the various shapes, which limits their clinical application. In this paper, we focus on utilizing multi-scale image contexts and attention mechanisms to improve networks’ ability to learn discriminative feature representation for accurate segmentation and present a novel FCNN architecture called multis-scale guided squeeze-and-excitation network (MSGSE-Net). In particular, we first propose the multi-scale guided squeeze-and-excitation (MSGSE) attention module which can progressively and selectively aggregate discriminative features. In contrast to existing attention modules, the MSGSE module performs an adaptive recalibration that features at different locations of the feature map are recalibrated under the guidance of multi-scale contexts. Then multi-scale spatial attention supervision is adopted to enhance the intra-class homogeneity and inter-class distinction of the attention weights. Moreover, we propose a novel entropy-weighted Dice loss (EDL) to force the network to focus on the ambiguous voxels around the boundaries of subcortical structures. We evaluate the proposed method on two challenging benchmark datasets (the IBSR dataset and the MALC dataset). The experimental results show that our model consistently yields better segmentation performance than several state-of-the-art methods and improves the segmentation Dice score by 1.6% at most compared with baseline method U-Net. Our code is available at https://github.com/neulxlx/MSGSE-Net.

Robust transfer learning based on Geometric Mean Metric Learning

Transfer learning usually utilizes the knowledge learned from the relative labeled source domain to promote the model performance in the unlabeled or few labeled target domain with different distribution. Most of the existing transfer learning methods aim to reduce the discrepancy of distributions between the source and target domains, but ignore the discriminative category information involved in the data from both domains in the process of knowledge transfer. To learn more discriminative feature representation in knowledge transfer, this paper integrates the transfer learning and metric learning into a unified framework and proposes a novel robust transfer learning based on geometric mean metric learning, namely Geometric Mean Transfer Learning (GMTL). GMTL uses weighted geometric mean metric learning to model the intra-class distance and the inter-class similarity. In the meantime, the marginal distributions and conditional distributions of the source and target domains are jointly adapted. Moreover, according to the natures of the datasets in different tasks, we dynamically combine the discriminative modeling and domain adaption to make the proposed model more robust. We assign different weights to the intra-class distance and the inter-class similarity in metric learning and different weights to marginal distribution adaption and conditional distribution adaption, respectively. Finally, the solution to the objective function is converted to the problem of finding a point on the geodesic joining two points on the Riemannian manifold, which is very simple and direct. Extensive experiments are conducted on six datasets widely adopted in transfer learning to verify the superiority of our proposed GMTL over existing state-of-the-art transfer learning methods.

Multiscale Feature Learning Based on Enhanced Feature Pyramid for Vehicle Detection

Vehicle detection is a crucial task in autonomous driving systems. Due to large variance of scales and heavy occlusion of vehicle in an image, this task is still a challenging problem. Recent vehicle detection methods typically exploit feature pyramid to detect vehicles at different scales. However, the drawbacks in the design prevent the multiscale features from being completely exploited. This paper introduces a feature pyramid architecture to address this problem. In the proposed architecture, an improving region proposal network is designed to generate intermediate feature maps which are then used to add more discriminative representations to feature maps generated by the backbone network, as well as improving the computational cost of the network. To generate more discriminative feature representations, this paper introduces multilayer enhancement module to reweight feature representations of feature maps generated by the backbone network to increase the discrimination of foreground objects and background regions in each feature map. In addition, an adaptive RoI pooling module is proposed to pool features from all pyramid levels for each proposal and fuse them for the detection network. Experimental results on the KITTI vehicle detection benchmark and the PASCAL VOC 2007 car dataset show that the proposed approach obtains better detection performance compared with recent methods on vehicle detection.

Learning auto-scale representations for person re-identification

Person re-identification (ReID) is a hot topic in computer vision. The data in the ReID is often collected from the cameras with different views and is affected by other environmental factors, which poses a significant challenge to ReID. The omni-scales proposed by OSNet can extract discriminative feature representations, which shows that omni-scales are adequate for the task of the ReID. However, the OSNet is mainly based on a manually designed network architecture. In the OSnet, each block uses the same architecture and has only four scale feature representations. Inspired by neural network architecture search (NAS), we propose a method of auto-scale representations for ReID. Specifically, we first design the auto-scale block, mainly composed of the Lite 3 × 3 operations and the RCB 1 × 1 operations. The connection status among the Lite 3 × 3 operations is just our search space. Then we give our entire macro network architecture, the auto-scale network, which is mainly composed of 6 auto-scale blocks. Unlike other NAS-related work, each block in our search space does not need to share the same architecture but can maintain a different architecture. In the search process, we propose the entropy regularization, the validity regularization and the consistent regularization to alleviate the discretized gap, no valid path, and meaningless edges, respectively. Finally, we verify the effectiveness of the model we searched on four commonly used datasets. Our model can maintain the same 2.2 M parameters as OSNet but can achieve the performance of SOTA. The mAP on the Market1501 dataset can reach 88.7%.

Design and Development of a Content-Based Image Retrieval (CBIR) System for Computing Similarity.

The shared and stored mixed media information is growing, and looking for or retrieving a significant image from a chronicle is a challenging exploration issue. Any image retrieval model's primary objective is to hunt for and mastermind photos that have a visual semantic association with the user's query. The bulk of web indexes on the Internet fetch photos using content-based algorithms that require subtitles as additional information. The user submits a query by inputting some text or keywords that match the file's keywords. )e yield is generated based on keyword matching, and this cycle can obtain insignificant photos. The distinction between human visual understanding and manual naming/commenting is the fundamental reason for producing the irrelevant yield. Any image retrieval framework must meet the fundamental criterion of searching for and sorting comparable photos from the archive with as little human interaction as possible. As implied by the writing, the choice of aesthetic characteristics for any framework is determined by the end user's requirements. Discriminative feature representation is another fundamental requirement for any image retrieval framework. To make the feature more robust and unique in terms of depiction fusion of low-level visual features, a large computational cost is required to obtain more dependable results. Regardless of the case, an ill-advised feature selection can degrade the performance of an image retrieval model. Contrary to conventional ideabased approaches, content-based picture retrieval is incompatible with them. "Content-based" refers to the fact that the hunt evaluates the image's contents rather than its metadata, such as keywords, labels, or depictions associated with the image.

Iterative graph attention memory network for cross-modal retrieval

How to eliminate the semantic gap between multi-modal data and effectively fuse multi-modal data is the key problem of cross-modal retrieval. The abstractness of semantics makes semantic representation one-sided. In order to obtain complementary semantic information for samples with the same semantics, we construct a local graph for each instance and utilize a graph feature extractor (GFE) to reconstruct the sample representation based on the adjacency relationship between the sample itself and its neighbors. Owing to the problem that some cross-modal methods only focus on the learning of paired samples and cannot integrate more cross-modal information from the other modalities, we propose a cross-modal graph attention strategy to generate the graph attention representation for each sample from the local graph of its corresponding paired sample. In order to eliminate heterogeneous gap between modalities, we fuse the features of the two modalities using a recurrent gated memory network to choose prominent features from other modalities and filter out unimportant information to obtain a more discriminative feature representation in the common latent space. Experiments on four benchmark datasets demonstrate the superiority of our proposed model compared with state-of-the-art cross-modal methods.

Multi-Scale Squeeze U-SegNet with Multi Global Attention for Brain MRI Segmentation.

In this paper, we propose a multi-scale feature extraction with novel attention-based convolutional learning using the U-SegNet architecture to achieve segmentation of brain tissue from a magnetic resonance image (MRI). Although convolutional neural networks (CNNs) show enormous growth in medical image segmentation, there are some drawbacks with the conventional CNN models. In particular, the conventional use of encoder-decoder approaches leads to the extraction of similar low-level features multiple times, causing redundant use of information. Moreover, due to inefficient modeling of long-range dependencies, each semantic class is likely to be associated with non-accurate discriminative feature representations, resulting in low accuracy of segmentation. The proposed global attention module refines the feature extraction and improves the representational power of the convolutional neural network. Moreover, the attention-based multi-scale fusion strategy can integrate local features with their corresponding global dependencies. The integration of fire modules in both the encoder and decoder paths can significantly reduce the computational complexity owing to fewer model parameters. The proposed method was evaluated on publicly accessible datasets for brain tissue segmentation. The experimental results show that our proposed model achieves segmentation accuracies of 94.81% for cerebrospinal fluid (CSF), 95.54% for gray matter (GM), and 96.33% for white matter (WM) with a noticeably reduced number of learnable parameters. Our study shows better segmentation performance, improving the prediction accuracy by 2.5% in terms of dice similarity index while achieving a 4.5 times reduction in the number of learnable parameters compared to previously developed U-SegNet based segmentation approaches. This demonstrates that the proposed approach can achieve reliable and precise automatic segmentation of brain MRI images.

Temporal Head Pose Estimation From Point Cloud in Naturalistic Driving Conditions

Head pose estimation is an important problem as it facilitates tasks such as gaze estimation and attention modeling. In the automotive context, head pose provides crucial information about the driver’s mental state, including drowsiness, distraction and attention. It can also be used for interaction with in-vehicle infotainment systems. While computer vision algorithms using RGB cameras are reliable in controlled environments, head pose estimation is a challenging problem in the car due to sudden illumination changes, occlusions and large head rotations that are common in a vehicle. These issues can be partially alleviated by using depth cameras. Head rotation trajectories are continuous with important temporal dependencies. Our study leverages this observation, proposing a novel temporal deep learning model for head pose estimation from point cloud. The approach extracts discriminative feature representation directly from point cloud data, leveraging the 3D spatial structure of the face. The frame-based representations are then combined with <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">bidirectional long short term memory</i> (BLSTM) layers. We train this model on the newly collected <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">multimodal driver monitoring</i> (MDM) dataset, achieving better results compared to non-temporal algorithms using point cloud data, and state-of-the-art models using RGB images. We further show quantitatively and qualitatively that incorporating temporal information provides large improvements not only in accuracy, but also in the smoothness of the predictions.

An ameliorated deep dense convolutional neural network for accurate recognition of casting defects in X-ray images

Recognizing defects in X-ray images plays an important role in the detection of internal defects in titanium alloy castings. However, the existing manual defects recognition methods have common drawbacks such as unstable artificial recognition, misrecognition, huge workload, and low efficiency of recognition. To make up for the shortcomings, an ameliorated deep dense convolutional neural network (BX-Net) was presented to accurately recognize casting defects in X-ray images and effectively extract highly discriminative features of different categories. DenseNet121 was used as the backbone of BX-Net and the feature extracted by DenseNet121 was fully shared by the two inputs of a bilinear pooling layer. Transfer learning was applied to reduce the demand for data and hyperparameters’ tuning. The backbone of BX-Net was firstly trained on the ImageNet and then all layers of BX-Net was fine-tuned on nine-hundred X-ray images of TiAl aero casting components. Other six deep convolutional neural networks (DenseNet121, EfficientNetB4, EfficientNetB7, ResNet50, VGG16 and Xception) were also trained to be compared with the presented BX-Net. Two Support Vector Machines were trained on the LBP features data set and HOG features data set of nine-hundred X-ray images of TiAl aero casting components respectively Experiments comparing BX-Net with other six deep learning models and two machine learning models on one hundred X-ray images (test set) of TiAl aero casting components were carried out. The comparison results show that BX-Net has the least parameters except for the Densenet121. The recall and accuracy of BX-Net were 99% and 99% respectively. In addition, the comparison results also show that BX-Net was the only model that learned discriminative feature representation of the casting X-ray image data set. The BX-Net proposed in this paper is expected to overcome the shortcomings of manual defects recognition.