Enterprise AI Analysis

Detection Head Optimization

YOLOv13 adopts three detection heads at 8×/16×/32× downsampling. While this design serves objects of different sizes, PCB defects are predominantly small; after 16×/32× downsampling they become too coarse, and deeper heads with large receptive fields introduce background noise. This multi-scale design exhibits significant mismatch issues in PCB defect detection scenarios. In contrast, the stride-8 (P3) feature map preserves higher resolution and is more suitable for precise localization of tiny defects. Based on this observation, we simplify the detector to focus on P3: we retain only the P3 head and remove P4/P5 together with their associated upsampling-fusion paths, concentrating computation on the most informative scale. Features from higher levels are kept only to provide semantic enhancement to P3 through streamlined fusion.

Dynamic Parameterized Gate Fusion Module (DPGF)

Feature fusion serves as a critical component in multi-scale object detection networks, integrating feature information from different levels. The FullPAD paradigm adopted by the original YOLOv13 balances feature contributions across different paths through learnable scalar parameters, expressed as: F = Fo + y. F₁ ... This paper proposes the Dynamic Pa- rameterized Gated Fusion (DPGF) module. This module achieves a more flexible feature aggregation mechanism through three tech- nical innovations. First, we adopt a unified parametric gating: a learnable scalar is mapped by a bounded Sigmoid into a gate, which weights the paths during fusion... Secondly, to address potential feature magnitude discrepancies across different network layers, DPGF performs RMS magnitude alignment on auxiliary features prior to fusion, scaling them to magnitudes comparable to the primary features... Finally, DPGF uses a path-aware initialization with differentiated gate settings. Lateral nodes use larger initial gates to encourage thorough fusion, whereas internal aggregation nodes adopt more conservative values. For small-scale feature paths (e.g., P3), initial gates are slightly increased to enhance small-object cues. The gate range is constrained by gmin and gmax to keep training stable, yielding reasonable feature weighting early in training and speeding up convergence.

Progressive Lightweight DSC3k2 Module (Lite-DSC3k2)

The DSC3k2 module serves as the core building block of YOLOv13 Neck, employing cross-channel branching and depth-separable convolutions for lightweight feature extraction. As shown in Figure 2(left), the original DSC3k2 consists of an outer 1×1 convolution layer and an internal DSC3k submodule. To further reduce model complexity, this paper proposes the lightweight Deep Separable Convolution module (Lite-DSC3k2) with a progressive grouped convolution strategy... The core innovation of Lite-DSC3k2 is its progressive grouped convolution strategy. As illustrated in Figure 2(right), this strat- egy employs decreasing group sizes across three nested layers: groups=4 for outer 1×1 convolutions, groups=2 for middle 1×1 convolutions, and standard convolution (g=1) for innermost point- wise convolutions. Grouped convolution is applied only to 1×1 convolutions, which account for the largest parameter share. This progressive design follows the "coarse-grained outside, fine-grained inside" principle: outer convolutions handle channel reduction and use larger group sizes; inner convolutions perform fine-grained extraction with greater impact, thus progressively reducing group sizes. Lite-DSC3k2 is applied to the Neck section of the YOLOv13-P3 ar- chitecture during two feature processing stages at different scales, handling 512-channel and 256-channel features respectively, as shown in Figure 1, both using the aforementioned configuration. Through the progressive grouping strategy, Lite-DSC3k2 achieves an optimal balance between lightweight design and performance, enhancing the model's deployability in resource-constrained environments.

Experimental Validation

Experimental results demonstrate that the proposed YOLOv13- PCB-Lite achieves good performance on the PCB defect dataset, with model parameters reduced by 39.2% and computational cost reduced by 22.6% compared to the baseline model, while maintain- ing comparable detection accuracy, with mAP50 decreasing by only 0.8%, proving the method's excellent balance between lightweighting and detection accuracy.

To validate the effectiveness of YOLOv13-PCB-Lite, a comprehen- sive comparative study was conducted with 10 representative meth- ods under identical dataset and evaluation protocols. The compared methods include traditional detectors (SSD[16], Faster R-CNN), Transformer-based methods (RT-DETR[17]), lightweight YOLO se- ries (SCF-YOLO[18], YOLOv5, YOLOv6[19], YOLOv7[20], YOLOv8, YOLOv10[21], YOLOv12[22]), and the original YOLOv13. Evalu- ation metrics encompass accuracy (mAP@0.5, Precision, Recall,) and efficiency (Params, GFLOPs). Detailed results are presented in Table 1.

To validate the contribution of each component, ablation stud- ies were conducted by progressively introducing three proposed improvements and comparing their performance impacts. Evaluation metrics encompass accuracy (Precision, Recall, mAP@0.5) and efficiency (Params, GFLOPs). Results are presented in Table 3.