GMBINet: Real-Time Steel Defect Detector

• The paper introduces GMBINet, a lightweight encoder-decoder network for real-time, pixel-level steel defect detection that achieves high accuracy with only 0.19M parameters.
• It employs a novel Group Multiscale Bidirectional Interactive (GMBI) module using bidirectional progressive interaction and parameter-free fusion to extract features efficiently.
• Experimental results demonstrate that GMBINet delivers competitive performance while being orders-of-magnitude more efficient than heavy models, making it ideal for edge device deployment.

GMBINet is a lightweight, real-time encoder–decoder network purpose-built for pixel-level detection of steel surface defects. It was introduced in “A Lightweight Group Multiscale Bidirectional Interactive Network for Real-Time Steel Surface Defect Detection” and is centered on a Group Multiscale Bidirectional Interactive (GMBI) module that performs multiscale feature extraction and cross-scale interaction while keeping computational cost invariant to the number of scales. At $512 \times 512$ input resolution, the reported model uses approximately $0.19$M parameters and $0.39$ GFLOPs, and achieves $1048$ FPS on GPU and $16.53$ FPS on CPU, with competitive accuracy on SD-Saliency-900 and NRSD-MN and additional validation on NEU-CLS (Zhang et al., 22 Aug 2025).

1. Industrial problem setting and design objective

Real-time surface defect detection is presented as a constraint-driven industrial vision problem: the model must operate on edge devices with limited compute and memory, while remaining effective on defects with large variation in scale, shape, and texture. The motivating claim is that heavy backbones such as ResNet, VGG, and Transformers provide strong multiscale semantics but incur high latency and high FLOPs and parameter counts, whereas existing lightweight methods often lose accuracy, especially for multiscale defects (Zhang et al., 22 Aug 2025).

The paper positions GMBINet against a specific weakness in lightweight multibranch depthwise separable convolution designs. In that family, complexity grows linearly with the number of branches because each branch processes the full tensor, and cross-scale interaction is weak because branches are isolated and typically fused only after feature extraction. GMBINet addresses both issues by moving interaction into the extraction process itself through channel grouping, bidirectional progressive interaction, and parameter-free fusion. This suggests that the architecture is not merely a smaller backbone, but a reformulation of how multiscale representations are computed under strict latency constraints.

The target output is a single-channel saliency or defect probability map at full image resolution. The stated deployment scenario includes CPU-only operation in resource-constrained industrial environments, which is important because the reported CPU throughput is substantially higher than that of heavy comparison models.

2. Network organization

The overall architecture is a 5-stage encoder–decoder. The encoder is a GMBI-based backbone, and the decoder uses progressive bilinear upsampling with depthwise separable convolution refinements and deep supervision at every decoder stage output $D_i$. All images are resized to $512 \times 512$ during training and inference, and the final prediction is restored to the original size by bilinear interpolation (Zhang et al., 22 Aug 2025).

Stage	Resolution / channels	Main operations
Stage 1	$256 \times 256$, 16 channels	Conv $3 \times 3$, stride 2
Stage 2	$128 \times 128$, 32 channels	DSConv stride 2, then $0.19$0 GMBI
Stage 3	$0.19$1, DSConv to 64 channels, GMBI $0.19$2	DSConv stride 2, then $0.19$3 GMBI
Stage 4	$0.19$4, 96 channels	DSConv stride 2, then $0.19$5 GMBI
Stage 5	$0.19$6, 128 channels	DSConv stride 2, then $0.19$7 GMBI

Stage 2 begins with DSConv $0.19$8, stride 2, expanding to 32 channels and reducing resolution from $0.19$9 to $0.39$0, followed by three GMBI blocks at $0.39$1. Stage 3 applies DSConv to 64 channels from $0.39$2 to $0.39$3, then four GMBI blocks at $0.39$4, with the note that the table lists 32 channels for the GMBI at this stage. Stage 4 applies DSConv to 96 channels from $0.39$5 to $0.39$6, then six GMBI blocks. Stage 5 applies DSConv to 128 channels from $0.39$7 to $0.39$8, then three GMBI blocks.

The decoder progressively upsamples and refines features using DSConv. Deep weighted supervision is applied at five decoder stages with equal weights, using a hybrid BCE + SSIM loss. The paper states that this stabilizes training and accelerates convergence. A plausible implication is that the decoder is intentionally simple so that most representational capacity remains in the GMBI backbone rather than being shifted into a heavy reconstruction head.

3. GMBI module, BPFI, and EWMS

The GMBI module is the defining unit of GMBINet. Its input tensor $0.39$9 is split into $1048$0 channel groups $1048$1, $1048$2. Each group is processed by a $1048$3 depthwise convolution with dilation $1048$4, so each group corresponds to a distinct receptive field. Because grouping makes the number of channels processed per scale equal to $1048$5, the total FLOPs are independent of the number of scales $1048$6 (Zhang et al., 22 Aug 2025).

The paper formalizes the depthwise separable convolution cost, under $1048$7, as

$1048$8

For GMBI, the multiscale extraction and fusion cost is

$1048$9

which is equal to a single DSConv block. By contrast, the multibranch DSConv baseline and the MI module both increase with $16.53$0.

The Bidirectional Progressive Feature Interactor (BPFI) introduces two-path interaction. Forward guidance uses bottom-up cues from smaller to larger scales. In the paper’s formulation,

$16.53$1

where $16.53$2 is the interaction operator. Backward enhancement then modulates smaller-scale features using larger-scale features; in the practical guidance, this is written as $16.53$3, and for $16.53$4, $16.53$5. The stated effect is refinement of representations while expanding effective receptive field, enhancing semantic consistency, and preserving details.

The interaction operator is the parameter-free Element-Wise Multiplication–Summation (EWMS):

$16.53$6

where $16.53$7 is sigmoid and $16.53$8 denotes element-wise multiplication. In the backward path, the higher-scale feature gates the lower-scale feature in the same manner. After bidirectional refinement, the outputs are concatenated, fused with a pointwise convolution, and added residually:

$16.53$9

This combination of group-wise multiscale extraction, BPFI, and residual pointwise fusion is the core technical claim of the model. The paper’s interpretation is that GMBI avoids both the linear complexity growth of multibranch designs and the weakness of post-hoc-only interactions.

4. Optimization protocol and benchmark configuration

Training uses an NVIDIA GeForce RTX 4090 (24 GB) and an Intel Core i9-13900KF (3.0 GHz), the Adam optimizer, an initial learning rate of $D_i$0, cosine annealing, batch size 32, and 50,000 iterations trained to convergence. The loss is a hybrid BCE + SSIM objective with deep weighted supervision:

$D_i$1

The augmentation pipeline consists of z-score normalization, random flipping, intensity shifts, and scaling (Zhang et al., 22 Aug 2025).

Evaluation is performed on three datasets. SD-Saliency-900 contains 900 grayscale images at $D_i$2, with three defect categories— inclusion, patches, and scratches—300 images each. Its training split is 810 images, comprising 540 defect images with 180 per type, plus 270 salt-and-pepper noise images. NRSD-MN contains 3,936 RGB images with varying resolutions, split into 2,086 train, 885 validation, and 965 test samples following MCNet. NEU-CLS contains 1,800 grayscale images from six categories, with 70% train, 10% validation, and 20% test.

The metrics on SD-Saliency-900 and NRSD-MN are MAE, WF, OR, SM, PFOM, IoU, and efficiency metrics including Params, FLOPs, and FPS. NEU-CLS uses accuracy, precision, recall, and F1. The choice of both saliency-style segmentation metrics and industrial efficiency metrics reflects the paper’s emphasis on simultaneous accuracy and deployment viability.

5. Quantitative performance and ablation structure

On SD-Saliency-900 at $D_i$3, GMBINet reports MAE $D_i$4, WF $D_i$5, OR $D_i$6, SM $D_i$7, PFOM $D_i$8, and IoU $D_i$9, with $512 \times 512$0M parameters, $512 \times 512$1G FLOPs, $512 \times 512$2 FPS on GPU, and $512 \times 512$3 FPS on CPU. On NRSD-MN, it reports MAE $512 \times 512$4, WF $512 \times 512$5, OR $512 \times 512$6, SM $512 \times 512$7, PFOM $512 \times 512$8, IoU $512 \times 512$9, again with $256 \times 256$0M parameters, $256 \times 256$1G FLOPs, and $256 \times 256$2 FPS (Zhang et al., 22 Aug 2025).

Against heavy baselines, DACNet reaches SM $256 \times 256$3 on SD-Saliency-900, which is $256 \times 256$4 higher than GMBINet, but requires $256 \times 256$5M parameters, $256 \times 256$6G FLOPs, $256 \times 256$7 FPS on GPU, and $256 \times 256$8 FPS on CPU. The paper summarizes this regime by stating that heavy backbones such as EDRNet, DACNet, EMINet, U2Net, EDN, and PoolNet are orders of magnitude heavier, while GMBINet matches or closely trails the top accuracy and is $256 \times 256$9–$3 \times 3$0 lighter and $3 \times 3$1–$3 \times 3$2 faster on GPU, with much higher CPU throughput.

Against lightweight competitors, the reported comparisons are more favorable. MINet yields MAE $3 \times 3$3, OR $3 \times 3$4, SM $3 \times 3$5, IoU $3 \times 3$6, with $3 \times 3$7M parameters, $3 \times 3$8G FLOPs, and $3 \times 3$9 FPS. BiSeNet reaches $128 \times 128$0 FPS but lower accuracy, including IoU $128 \times 128$1. CSNet uses $128 \times 128$2M parameters but $128 \times 128$3G FLOPs, and is both less accurate and slower. The paper states that GMBINet leads on all accuracy metrics against lightweight competitors while having the lowest FLOPs, nearly the smallest parameter count, and the fastest throughput.

The ablation studies isolate the contribution of each design choice. Removing multiscale processing entirely gives MAE $128 \times 128$4, SM $128 \times 128$5, and $128 \times 128$6 FPS. A classic branch-based multiscale design reaches MAE $128 \times 128$7, SM $128 \times 128$8, but increases cost to $128 \times 128$9M parameters, $0.19$00G FLOPs, and lowers speed to $0.19$01 FPS. GMBI reaches MAE $0.19$02, SM $0.19$03, with $0.19$04M parameters, $0.19$05G FLOPs, and $0.19$06 FPS. For BPFI, removing interaction yields MAE $0.19$07, SM $0.19$08, and $0.19$09 FPS; using forward guidance only gives MAE $0.19$10, SM $0.19$11; using backward enhancement only gives MAE $0.19$12, SM $0.19$13; using both yields the final MAE $0.19$14, SM $0.19$15, and $0.19$16 FPS. The paper’s interpretation is explicit: forward and backward paths are complementary and yield substantial accuracy gains with negligible speed loss.

For interaction type, EWMS gives the best overall accuracy without increasing parameters. Sum yields MAE $0.19$17, WF $0.19$18, OR $0.19$19, SM $0.19$20, PFOM $0.19$21, and $0.19$22 FPS. Multiply yields MAE $0.19$23, SM $0.19$24, and $0.19$25 FPS. Concatenation reaches MAE $0.19$26, WF $0.19$27, OR $0.19$28, SM $0.19$29, PFOM $0.19$30, but requires $0.19$31M parameters and $0.19$32G FLOPs and slows to $0.19$33 FPS.

The scale parameter $0.19$34 is also ablated. All tested values preserve $0.19$35M parameters and $0.19$36G FLOPs, while FPS decreases mildly with larger $0.19$37: $0.19$38 FPS at $0.19$39, $0.19$40 FPS at $0.19$41, and $0.19$42 FPS at $0.19$43. Accuracy peaks at $0.19$44, where MAE, WF, OR, SM, PFOM, and IoU reach $0.19$45, $0.19$46, $0.19$47, $0.19$48, $0.19$49, and $0.19$50, respectively. The paper attributes degradation at very large $0.19$51 to diminished per-group channel capacity.

6. Generalization, deployment, limitations, and nomenclature

A GMBI-based backbone classifier on NEU-CLS reports accuracy $0.19$52, precision $0.19$53, recall $0.19$54, F1 $0.19$55, $0.19$56M parameters, $0.19$57G FLOPs, and $0.19$58 FPS. ResNet50 reaches accuracy $0.19$59 but uses $0.19$60M parameters, $0.19$61G FLOPs, and $0.19$62 FPS. The paper states that the GMBI classifier achieves more than $0.19$63 speed and approximately $0.19$64 of the parameters with a small accuracy drop, and uses this to argue for broader industrial vision applicability beyond surface defect detection (Zhang et al., 22 Aug 2025).

The qualitative analysis states that on SD-Saliency-900, GMBINet avoids over-segmentation and under-segmentation seen in PoolNet, U2Net, EDRNet, SegFormer, and SeaFormer, and preserves connectivity where MINet exhibits discontinuities. On NRSD-MN, it aligns better with human annotations than CSNet and MINet, which over-segment, and U2Net and CGNet, which under-segment. The paper attributes this behavior to enhanced cross-scale communication.

The stated limitations are also specific. Accuracy can still lag top heavy models marginally, as illustrated by SM $0.19$65 versus $0.19$66. Extremely large $0.19$67 degrades performance because of over-splitting channels. Challenging conditions such as poor illumination and reflective surfaces may affect robustness. Proposed future directions are multimodal fusion with depth or thermal data, and hardware-aware optimization through pruning, distillation, and quantization.

For deployment, the paper recommends $0.19$68 for the best accuracy–speed balance, keeping input at $0.19$69 for the published performance, and considering integer-factor resizing pipelines to preserve cache locality. It further proposes quantization-aware training and pruning for microcontroller-class targets, and suggests replacing residual blocks in classification backbones with GMBI units, or plugging GMBI into encoder stages for segmentation tasks on other materials such as glass, aluminum, PCB, and rails.

A recurring source of confusion is nomenclature. In the steel-defect literature, GMBINet refers to the Group Multiscale Bidirectional Interactive Network introduced in 2025 (Zhang et al., 22 Aug 2025). However, nearby acronyms have been used for unrelated models: GM-Net denotes “Grouped Merging Net” (Chen et al., 2017); BI-GreenNet denotes a boundary-integral network for Green’s functions (Lin et al., 2022); GreenMGNet denotes “Green Multigrid Network” (Lin et al., 2024); and Multi-Grid Back-Projection Networks have been loosely associated with similar shorthand in challenge reports (Michelini et al., 2021). The practical implication is that GMBINet should be identified by its full title and application domain rather than by acronym alone.

Code and data are reported as publicly available at the project repository linked in the paper.