Conveyor-Belt Operational Status Monitoring

This repository contains our end-to-end information-fusion framework based on Mask R-CNN for joint detection and segmentation of conveyor-belt operating states. We integrate three novel modulesMS-DFF, TOM, and BEsegalong with a Dynamic Weighted Hybrid Loss to deliver superior performance in challenging industrial scenarios with widely varying object scales.

Table of Contents

• Features
  
• Installation
  
• Dataset
  
• Usage
  
• Results
  
• Citation
  

Features

• Multi-Scale Dynamic Feature Fusion (MS-DFF)
  Aggregates and adapts pyramid features for both detection and segmentation in one unified backbone.

• Task-Oriented Module (TOM)
  Dynamically routes fused features to the appropriate task head to improve multi-task collaboration.

• Boundary Enhanced Segmentation (BEseg)
  Reinforces mask boundaries for clearer, more accurate segmentation.

• Dynamic Weighted Hybrid Loss (DWH Loss)
  Automatically balances detection and segmentation losses during training.

Installation

1. Clone this repository:
   bash git clone https://github.com/Ashores/TOMSDFF.git cd TOMSDFF/

2. Create a Python environment and install dependencies:

bash python3 -m venv venv source venv/bin/activate pip install -r requirements.txt

Dataset

• Belt Dataset

  • Annotated in COCO format
  • Publicly available at: data/
  • Contains images and JSON annotation files for conveyor-belt anomalies

Usage

1. Prepare the dataset following COCO directory structure:

data/ train2017/ val2017/ annotations/ instances_train2017.json instances_val2017.json 2. Train the model:

bash python train.py --config configs/ours_mask_rcnn.yaml 3. Evaluate on the validation set:

bash python evaluate.py --config configs/ours_mask_rcnn.yaml 4. Visualize results:

bash python demo.py --config configs/ours_mask_rcnn.yaml --input path/to/image.jpg

Results

| Model | AP (%) | AP (%) | mAP (%) | AP (%) (Seg) | AP (%) (Seg) | mAP (%) (Seg) | FLOPs (Gb) | Params (Mb) | | ------------- | ---------- | ---------- | --------- | ---------------- | ---------------- | --------------- | ---------- | ----------- | | Mask-R-CNN | 97.3 | 64.7 | 66.0 | 54.4 | 40.2 | 39.7 | 142 | 43.99 | | Cascade-R-CNN | 97.9 | 65.6 | 67.2 | 55.2 | 41.5 | 42.4 | 163 | 77.03 | | SOLO | 97.5 | 68.3 | 67.3 | 63.4 | 42.2 | 43.1 | 156 | 36.13 | | SOLOv2 | 97.9 | 66.4 | 68.0 | 70.1 | 43.6 | 42.7 | 139 | 46.24 | | HTC | 97.7 | 67.1 | 67.6 | 56.3 | 42.1 | 41.5 | 162 | 79.97 | | Yolact | 98.0 | 66.2 | 66.9 | 60.6 | 43.5 | 42.3 | 61.7 | 34.75 | | DynaMask | 97.8 | 67.7 | 67.8 | 58.7 | 43.1 | 42.8 | 782 | 56.3 | | OrientedR-CNN | 98.1 | 67.6 | 67.1 | 57.9 | 42.7 | 42.1 | 136 | 51.7 | | YOLO11-seg | 98.3 | 67.9 | 68.7 | 65.5 | 44.2 | 43.6 | 123.3 | 22.4 | | Ours | 98.4 | 70.2 | 69.8 | 73.5 | 46.8 | 45.8 | 339 | 63.72 |

Our approach achieves the best overall detection and segmentation performance while maintaining a reasonable computational footprint.

Citation