Research on Non-Pooling YOLOv5 Based Algorithm for the Recognition of Randomly Distributed Multiple Types of Parts

Zehua Yu, Ling Zhang, Xingyu Gao, Yang Huang, Xiaoke LiuSensors 2022 · DOI

TL;DR

This paper introduces a "non-pooling" YOLOv5 (NP-YOLOv5) architecture to improve the recognition of randomly distributed industrial parts, a common challenge in automated assembly. By replacing pooling layers with convolutional layers and using the Mish activation function, their method achieves higher precision and recall rates compared to traditional YOLOv5, making robotic part handling more efficient.

What Problem Does This Paper Address?

Precision machinery assembly relies on clean parts, but after cleaning, these parts are often scattered randomly in collection bins. This poses a significant challenge for robotic systems, which typically struggle with identifying and classifying objects in unpredictable positions. Existing robots are often limited to picking parts from fixed locations, and vision-guided systems can falter when dealing with multiple part types mixed together. Manual collection is labor-intensive and costly. This paper aims to bridge this gap by developing a robust and efficient method for recognizing and classifying these randomly distributed, multi-type parts using computer vision.

Methods

The core innovation is the development of a "non-pooling" YOLOv5 (NP-YOLOv5) network. Traditional YOLOv5, like many deep learning architectures, uses pooling layers (e.g., in the Spatial Pyramid Pooling - SPP module) to downsample feature maps. While effective for capturing multi-scale information, pooling layers can lead to information loss and pose challenges for backpropagation, especially in multi-scale pooling scenarios.

The authors propose replacing these pooling layers with convolutional layers, specifically introducing a Spatial Pyramid Convolution (SPC) structure. This SPC module effectively replaces the SPP module in YOLOv5, ensuring that the entire network consists only of convolutional layers. This "non-pooling" approach aims to retain more spatial information and improve the gradient flow during training.

Additionally, the paper investigates the impact of different activation functions. While traditional YOLOv5 uses SiLU, the authors explore and ultimately select the Mish activation function. Mish is a non-saturating function that is smoother than ReLU and is shown to be more conducive to gradient computation and updating, leading to better accuracy and noise resistance, especially in larger networks.

Datasets

The paper utilizes self-created datasets of industrial parts for training and evaluation. The specific details of these datasets, such as the number of part types, images, or their specific characteristics, are not extensively detailed beyond being "self-created datasets of parts."

Key Insights and Ideas

  • Pooling Layers are a Bottleneck: The paper highlights that pooling layers, while common, can lead to information loss and complicate training, especially in complex multi-scale scenarios like SPP.
  • Convolutional Replacements are Viable: Replacing pooling with convolutional layers (SPC) is a promising strategy to retain spatial information and improve network performance.
  • Mish Activation Function's Superiority: The Mish activation function demonstrates a clear advantage over SiLU and ReLU in this context, leading to higher accuracy and better training stability. Its non-monotonic nature and smoothness contribute to improved gradient flow.
  • Synergy of NP-YOLOv5 and Mish: The combination of the non-pooling architecture and the Mish activation function yields the best performance, suggesting a synergistic effect.

Key Results

The proposed NP-YOLOv5 network, particularly when combined with the Mish activation function (Mish-NP), demonstrated significant improvements over traditional YOLOv5:

  • Precision Improvement: The NP-YOLOv5 network improved part recognition precision by 8% within 100 epochs of training.
  • Recall Rate Improvement: The NP-YOLOv5 network improved part recall rate by 3% within 100 epochs of training.
  • Overall Performance: After 100 epochs, the Mish-NP network achieved a precision of 0.96 and a recall of 0.96, outperforming other tested methods.
  • mAP Scores: The Mish-NP network achieved the highest average mAP across various IOU thresholds (0.5:0.05:0.95), reaching 0.84, compared to SiLU-YOLOv5 (0.837) and AM-YOLOv5 (0.828).
  • Faster Convergence: The Mish-NP network showed a faster convergence rate for [email protected], reaching above 0.95 after 110 epochs, compared to traditional YOLOv5 variants.
  • Higher Confidence Scores: Visual inspection of detection results indicated that the Mish activation function yielded higher part recognition confidence scores than SiLU.

Limitations and Future Work

  • Dataset Specificity: The authors acknowledge that the datasets used are self-created. They plan to investigate the applicability of their method to more diverse datasets in the future.
  • Limited Network Comparison: The research primarily focused on replacing pooling in YOLOv5. The paper did not explore whether similar convolutional replacements would yield benefits in other network architectures.
  • Future Work: The authors suggest exploring the effectiveness of their method on a wider range of datasets and potentially applying it to other network architectures to further validate its generalizability.

Summary updated

Abstract

Part cleaning is very important for the assembly of precision machinery. After cleaning, the parts are randomly distributed in the collection area, which makes it difficult for a robot to collect them. Common robots can only collect parts located in relatively fixed positions, and it is difficult to adapt these robots to collect at randomly distributed positions. Therefore, a rapid part classification method based on a non-pooling YOLOv5 network for the recognition of randomly distributed multiple types of parts is proposed in this paper; this method classifies parts from their two-dimensional images obtained using industrial cameras. We compared the traditional and non-pooling YOLOv5 networks under different activation functions. Experimental results showed that the non-pooling YOLOv5 network improved part recognition precision by 8% and part recall rate by 3% within 100 epochs of training, which helped improve the part classification efficiency. The experiment showed that the non-pooling YOLOv5 network exhibited improved classification of industrial parts compared to the traditional YOLOv5 network.
BibTeX
@inproceedings{yu2022research,
  author = {Zehua Yu and Ling Zhang and Xingyu Gao and Yang Huang and Xiaoke Liu},
  title = {{Research on Non-Pooling YOLOv5 Based Algorithm for the Recognition of Randomly Distributed Multiple Types of Parts}},
  booktitle = {Sensors},
  year = {2022},
  doi = {10.3390/s22239335},
}
Figure