C2f-EMCM Module: Efficient Multi-Scale Extraction

• C2f-EMCM module is an advanced convolution design that integrates dual-scale (3×3 and 5×5) operations to optimize feature extraction in lightweight models.
• It employs channel grouping and parallel multi-scale convolutions to address redundant computations and improve feature representation.
• Empirical results show up to a 3% increase in mAP and a 32.3% reduction in parameters, enabling efficient real-time edge deployments.

The C2f-EMCM module is an architectural enhancement for lightweight convolutional neural networks, specifically developed to balance detection accuracy and computational efficiency in real-time image analysis tasks. It integrates within the YOLOv8n object detection framework and is designed to optimize feature extraction through a dual strategy: multi-scale convolution and redundancy minimization. The C2f-EMCM module addresses the challenges associated with extracting features of varying spatial scales—essential in domains such as shrimp disease detection—while reducing the model’s parameter count and computational load (2507.02354).

1. Architectural Motivation

YOLOv8n's baseline C2f module utilizes standard 3×3 convolutions within a cross-stage partial residual structure. However, two limitations are evident:

• Limited Multi-scale Feature Extraction: Fixed-size (3×3) convolutions constrain the network's ability to capture features from targets of different sizes and lesion patterns.
• Redundant Computation: Recurrent application of 3×3 convolutions introduces computational redundancy, particularly within multi-branch bottleneck stages, elevating cost without proportional accuracy gains.

The C2f-EMCM module is engineered to overcome these issues by augmenting the C2f block with an Efficient Multi-scale Convolution Module (EMCM), thus improving both scale-diverse feature extraction and computation efficiency.

2. Module Structure and Operation

The C2f-EMCM module is designed as follows:

• Channel Grouping: The input feature map is split across the channel dimension into two groups:
  • The first group preserves the “original feature,” acting as a shortcut path to retain initial information.
  • The second group is subjected to multi-scale processing.
• Parallel Multi-scale Convolutions:
  • One branch applies a 3×3 convolution, favoring fine texture and edge extraction.
  • The other branch applies a 5×5 convolution, capturing coarser spatial context.
• Feature Concatenation and Mixing:
  • Output from both convolution paths is concatenated with the preserved channels.
  • A 1×1 convolution is then used to mix channels, promote cross-feature interaction, and adapt the output dimensionality.

The workflow can be represented mathematically as: $$\text{Out} = \text{Conv}_{1\times 1} \Big( \text{Concat}(\text{Feature}_{\text{original}}, \text{Conv}_{3\times 3}(\text{Feature}_{\text{split}}), \text{Conv}_{5\times 5}(\text{Feature}_{\text{split}})) \Big)$$

The repeated 3×3 convolutions of the original C2f are thus replaced with this dual-path structure, leading to more efficient and expressive feature extraction.

3. Bottleneck-EMCM Integration

The EMCM is further embedded within a modified bottleneck block—referred to as Bottleneck-EMCM—which fully replaces the standard convolutional operations of the C2f bottleneck. The updated C2f-EMCM replaces the last two backbone and two neck C2f modules in YOLOv8n, directly impacting both shallow and deep feature transformation throughout the network.

4. Impact on Computational Efficiency and Accuracy

The C2f-EMCM module yields several measurable improvements:

• Multi-scale Feature Extraction: The explicit use of both 3×3 and 5×5 paths facilitates optimized representation for small and large lesions—critical for accurate shrimp disease detection.
• Reduced Computational Complexity: By restricting the heavier multi-scale operations to only part of the feature map (via channel grouping), overall MACs and parameter counts are reduced compared to applying large convolutions globally.
• Improved Feature Fusion: The 1×1 convolution consolidates information and mitigates redundancy, enabling deeper network deployment with less risk of overfitting or unnecessary computation.

5. Integration Within the YOLOv8n Framework

The enhanced C2f-EMCM modules seamlessly replace selected C2f blocks in both the backbone and neck of YOLOv8n. This enables:

• Maintenance or improvement of detection accuracy through more expressive representation.
• A lighter model suitable for real-time edge deployment due to substantial reduction in parameter count and computational cost.

6. Empirical Validation and Metrics

Empirical results using a custom shrimp disease dataset demonstrate the C2f-EMCM module’s efficacy:

• When the EMCM is individually introduced, YOLOv8n’s precision increases to 86.5% (from 78.4%), and [email protected] rises by 2.4% to 92.1%.
• In the final model, employing C2f-EMCM in combination with other modules yields a [email protected] of 92.7%, exceeding the baseline YOLOv8n by 3%.
• The number of parameters drops by 32.3% (from 3.1M to 2.1M), making the architecture notably more efficient.
• These improvements persist in cross-dataset generalization experiments (e.g., a 4.1% [email protected] increase on the URPC2020), supporting the module’s robustness (2507.02354).

7. Practical Implications and Deployment

The C2f-EMCM module demonstrates that judicious architectural redesign—channel-splitting, parallel multi-scale convolutions, and final channel mixing—can significantly improve both detection accuracy and computational efficiency. This is particularly relevant for applications in aquaculture, medical imaging, and other domains where high performance must be attained under strict resource constraints.

Systems employing C2f-EMCM are well-suited to real-time edge deployment scenarios, supporting high-throughput and energy-efficient inference without loss in detection reliability. Its integration within YOLOv8n sets a precedent for similar multi-scale and lightweight module designs in other object detection frameworks that target embedded or real-time vision tasks.