PaperlayerView exact revision on arXiv

Source-linked AI summary

LF-YOLO: A Lighter and Faster YOLO for Weld Defect Detection of X-ray Image

Moyun Liu, Youping Chen, Lei He, Yang Zhang, Jingming Xie

arXiv:2110.15045v2cs.CV

TL;DR

X-ray weld-defect detection is challenging because defect shapes and scales vary substantially. LF-YOLO addresses this with RMF for combined parameter-based and parameter-free multi-scale extraction and EFE for efficient feature processing, reaching 92.9 mAP50 at 61.5 FPS on the weld-defect dataset and 50.3 mAP50 on MS COCO.

  • Problem

    Weld defects vary in scale, shape, aspect ratio, and contextual appearance, requiring models to capture both local and global information.

  • Method

    LF-YOLO combines an EFE module for low-consumption feature extraction with an RMF module that fuses parameter-based and parameter-free multi-scale information.

  • Results

    92.9 mAP50 at 61.5 FPS with 4.0G FLOPs and 7.3M parameters was achieved on the weld-defect dataset; the ×1.25 model reached 50.3 mAP50 on MS COCO.

  • Takeaways & Limitations

    LF-YOLO achieves a balance between detection performance and computational consumption while remaining competitive on MS COCO.

Abstract

from arXiv · show

X-ray image plays an important role in manufacturing industry for quality assurance, because it can reflect the internal condition of weld region. However, the shape and scale of different defect types vary greatly, which makes it challenging for model to detect weld defects. In this paper, we propose a weld defect detection method based on convolution neural network, namely Lighter and Faster YOLO (LF-YOLO). In particularly, a reinforced multiscale feature (RMF) module is designed to implement both parameter-based and parameter-free multi-scale information extracting operation. RMF enables the extracted feature map capable to represent more plentiful information, which is achieved by superior hierarchical fusion structure. To improve the performance of detection network, we propose an efficient feature extraction (EFE) module. EFE processes input data with extremely low consumption, and improves the practicability of whole network in actual industry. Experimental results show that our weld defect detection network achieves satisfactory balance between performance and consumption, and reaches 92.9 mean average precision mAP50 with 61.5 frames per second (FPS). To further prove the ability of our method, we test it on public dataset MS COCO, and the results show that our LF-YOLO has a outstanding versatility detection performance. The code is available at https://github.com/lmomoy/LF-YOLO.

I. INTRODUCTION

X-ray imaging supports weld quality inspection, but defect detection is difficult because defects vary in scale and shape, boundaries are blurred, and foreground and background have similar textures. LF-YOLO addresses these challenges with RMF and EFE modules designed for multi-scale representation and efficient feature extraction.

  • Manual weld inspection is inefficient and unsustainable because workers are subjective and easily fatigued.
  • Weld defects vary greatly in scale and shape, while blurred boundaries and similar foreground-background textures hinder recognition.
  • Prior weld-defect studies rarely adapt CNNs to the scale variation and contextual-information requirements of weld images.
  • RMF combines parameter-based and parameter-free operations to fuse local and global cues for richer multi-scale features.
  • EFE forms the LF-YOLO backbone and extracts meaningful features with few parameters and low computation.
  • LF-YOLO is proposed as a memory- and computation-friendly detector for multiple X-ray weld-defect classes and is evaluated on proprietary and public datasets.

II. RELATED WORKS

CNN-based object detectors use one-stage or two-stage strategies, while lightweight networks reduce the computation that limits industrial deployment. Representative lightweight designs reduce parameters through operations such as 1×1 convolution, depthwise separable convolution, or cheap feature generation.

  • Object detection networks are mainly divided into one-stage and two-stage categories.
  • One-stage detectors directly regress target classification and location, enabling fast detection and hardware-friendly operation.
  • High CNN computation requirements limit industrial application, motivating lightweight networks that reduce model complexity.
  • SqueezeNet, MobileNet, and GhostNet reduce computation or parameters using 1×1 convolution, depthwise separable convolution, or cheap operations.

B. Multi-scale feature utilization

CNN features are scale-sensitive, so multi-scale methods combine information from different receptive fields or existing network features. These approaches motivate the paper’s use of efficient multi-scale feature utilization for weld-defect detection.

  • CNN convolutional features are scale-sensitive, especially when object sizes vary greatly.
  • Parameter-based methods learn implicit multi-scale features using operations such as dilated convolutions, ASPP, and varied kernel sizes.
  • Parameter-free methods reuse existing features without introducing new learnable units, combining feature maps across layers.
  • FPN and PANet propagate semantic and spatial information through top-down and bottom-up paths, while spatial pyramid pooling provides another parameter-free strategy.
  • Prior weld-defect studies applied neural networks and CNNs, but the cited work provides the context for LF-YOLO’s multi-scale detector design.

A. EFE module

The EFE module is a lightweight feature-extraction block for X-ray weld images that expands, transforms, and merges features while reducing computation. It combines split-transform-merge processing with Ghost Conv to balance representation learning and complexity.

  • A. EFE module: EFE is designed as a lightweight, high-performance representation-learning module for X-ray weld images.Feature extraction is a major source of network parameters and computation, so the module targets whole-network efficiency.
  • A. EFE module: The module first uses 1×1 convolution to expand input channels to 2c or maintain c, depending on its network location.The input feature map has dimensions h×w×c.
  • A. EFE module: After convolution, EFE splits features into an identity branch and a dense feature-extraction branch using a split ratio of 0.25.The identity branch reuses input data, while the other branch further extracts features.
  • A. EFE module: Ghost Conv generates intrinsic feature maps with conventional convolution and produces additional ghost maps through cheap linear operations.This reduces redundant computation while producing the required output channel number.
  • A. EFE module: The two branches are concatenated, then a second 1×1 convolution compresses channels before residual element-wise addition with the expansion input.The expansion output dimension is set to twice the input, and the residual structure supports information flow across layers.
  • A. EFE module: Figure 3 presents EFE as channel-space expansion with split-transform-merge reuse and Ghost Conv-based complexity reduction.The module is intended to encode more implicit information while limiting computational cost.

B. RMF module

The RMF module addresses scale sensitivity by hierarchically combining parameter-free pooling with parameter-based GDConv features from different receptive fields.

  • B. RMF module: RMF combines parameter-free maxpooling with parameter-based GDConv to obtain multiscale contextual information.The maxpooling stage forms a base pyramid, while GDConv learns additional features within each branch.
  • B. RMF module: Different maxpool sizes generate explicit multiscale feature maps before subsequent learnable processing.This stage exploits existing feature-map information without introducing parameters.
  • B. RMF module: GDConv retains Ghost Conv’s structure while applying depthwise convolutions with dilation rates of 1, 5, and 9.These branches capture cues from varied receptive fields with a lightweight design.
  • B. RMF module: The outputs from RMF’s hierarchical branches are concatenated to integrate explicit and implicit multiscale information for finer prediction.The design uses three GDConvs and four maxpool branches.

C. The architecture of LF-YOLO

LF-YOLO combines EFE, maxpooling, RMF enhancement, and three scale-specific prediction heads in a lightweight detection architecture.

  • C. The architecture of LF-YOLO: LF-YOLO uses CBL for initial extraction, maxpooling for resolution reduction, and EFE for deeper representation learning.The backbone begins from an H×W ×3 weld image input.
  • C. The architecture of LF-YOLO: Table II presents the backbone network used by LF-YOLO.The supplied passage identifies the backbone table but does not provide its individual rows.
  • C. The architecture of LF-YOLO: RMF enhances backbone features before three detection heads predict objects at different sizes.The network inherits YOLO’s feature pyramid structure and adjusts head features using 1×1 CBL and Ghost Conv.
  • C. The architecture of LF-YOLO: LF-YOLO computes objectness and classification loss with binary cross entropy and regresses bounding boxes using IOU loss.The total loss combines these three components.

IV. EXPERIMENTS

The experiments compare LF-YOLO with state-of-the-art detectors, assess EFE and RMF through ablations, and evaluate versatility on MS COCO.

  • IV. EXPERIMENTS: The study compares LF-YOLO with state-of-the-art methods on a weld defect dataset.The experimental section also includes ablation studies for EFE and RMF.
  • IV. EXPERIMENTS: Ablation studies evaluate the contributions of the EFE and RMF modules.The experiments additionally compare classical models on a public dataset to assess versatility.
  • IV. EXPERIMENTS: The network’s versatility is evaluated through quantitative comparisons with classical models on MS COCO.The supplied passage describes this evaluation purpose without reporting its results.

A. Experimental Setup

The experiments use a professionally annotated weld-defect dataset, specified training settings, and mAP50, FLOPs, Params, and FPS for evaluation.

  • A. Experimental Setup: The weld dataset contains 2449 digital X-ray images, three defect categories, and a 9:1 train-test split.The categories are blow hole, incomplete penetration, and crack, with professional annotations.
  • A. Experimental Setup: MS COCO is added as a public dataset to evaluate LF-YOLO’s versatility.The supplied setup passage identifies MS COCO’s 118k training images and 80 categories.
  • A. Experimental Setup: Training uses PyTorch, SGD with 0.9 momentum and 0.0005 weight decay, and 500 or 300 epochs for the weld or MS COCO datasets.Images are resized to 320×320 for the weld dataset and 640 for MS COCO.
  • A. Experimental Setup: The evaluation reports mAP50, FLOPs, Params, and FPS to measure detection ability, computation, model size, and inference speed.FPS is averaged over 245 weld-dataset testing images.
  • A. Experimental Setup: Table III contains weld-defect detection results comparing LF-YOLO with state-of-the-art methods.The supplied table passage provides the table title but not its numerical entries.

B. Comparisons with state-of-the-art models

LF-YOLO outperforms the compared detection models on the weld-defect dataset while using substantially fewer computational resources. Its per-class predictions remain strong across defect types, although blow-hole detection is comparatively weaker.

  • 92.9 mAP50 is achieved by LF-YOLO, exceeding all compared methods on the weld-defect dataset.The model uses 7.3M Params and 4.0G FLOPs.
  • 7.3M Params and 4.0G FLOPs give LF-YOLO lower complexity than the compared classical one-stage and two-stage models.
  • LF-YOLO×0.5 achieves the best detection speed and parameter count while outperforming all lightweight networks, most one-stage models, and some two-stage networks.
  • LF-YOLO predicts all defect classes strongly, but blow-hole detection is weaker because its labels are sometimes dense and extremely small.

C. Ablation studies

Ablation studies show that the complete EFE and RMF designs provide the strongest efficiency–accuracy balance. Removing either RMF component substantially reduces model capacity, while the public-dataset test evaluates broader versatility.

  • EFE ablation: The complete EFE module provides the best balance of detection performance, computation, and storage.Replacing Ghost Conv with normal Conv raises mAP50 to 90.8 but increases FLOPs 2.7× and Params 2.3×.
  • EFE ablation: 90.8 mAP50 is obtained with normal Conv, only 0.1 point above the intact EFE module but at 2.7× FLOPs and 2.3× Params.
  • RMF ablation: RMF performs satisfactorily only when SPP and GDConv are used together; removing either parameter-free or parameter-based processing decreases model capacity.
  • RMF ablation: 86.3 mAP50 results when both RMF components are removed, representing the worst ablation degeneration.
  • Public-dataset validation: On MS COCO, LF-YOLO×0.75 reaches 43.9 mAP50 with computation equal to 28.7% and 25.9% of SSD300 and Faster-RCNN, respectively.LF-YOLO×1.25 performs better than most compared models with lower FLOPs and fewer Params, while R-FCN reaches 52.9 mAP50.

E. Visualization of intermediate feature maps

Visualization indicates that EFE extracts region divisions, fine-grained brightness variations, texture information, and geometrical cues from X-ray weld images. These representations support the paper’s low-consumption feature-extraction design.

  • Feature-map visualization: EFE feature maps distinguish different regions and represent fine-grained brightness variations that are difficult to notice in the original image.
  • Feature-map visualization: Channels 9, 24, and 32 emphasize underlying texture information, while channels 8, 11, and 31 focus more purely on region segmentation.
  • Feature-map visualization: EFE encodes region distribution and geometrical cues while using low consumption as the basic feature-extraction block.
  • Multi-scale representation: RMF combines parameter-free multi-scale information with parameter-based learning across receptive fields before fusing the results into a stronger representation.
  • Reported performance: LF-YOLO reaches 92.9 mAP50, 4.0G FLOPs, 7.3M Params, and 61.5 FPS on the weld-defect dataset, while its ×1.25 version reaches 50.3 mAP50 on MS COCO.
Loading 2110.15045v2…