Method for intelligent sorting, detection, and recognition of construction waste

Smart overview of the Invention

The method focuses on the intelligent sorting, detection, and recognition of construction waste using advanced computer vision techniques. By employing an improved SRGAN algorithm for preprocessing and an enhanced YOLOv8 model for detection, the approach aims to replace manual labor in sorting construction waste. This innovation addresses issues such as image feature loss due to conveyor belt vibrations and mutual occlusion of waste materials, which are common in traditional methods.

Technical Background

The rapid urbanization and increase in construction activities have led to significant amounts of construction waste, often disposed of in landfills without proper recycling. Many materials, such as stones, plastics, and wood, can be reused or recycled. Traditional sorting methods involve mechanical and manual operations, which are inefficient and pose health risks due to dust and noise. The need for a more efficient method using computer vision has become critical.

Problem Addressed

Current sorting methods suffer from low recycling purity and inefficiency, particularly in dusty environments where image features can become blurred. The new method aims to overcome these challenges by improving the accuracy of detection and recognition processes. It specifically targets the problems caused by conveyor belt vibrations and material occlusions that lead to feature loss in images.

Methodology

The process begins with collecting construction waste images at sorting sites. These images are preprocessed using an improved SRGAN algorithm and divided into training, validation, and test sets. The enhanced YOLOv8 model incorporates receptive field attention convolutions and multidimensional collaborative attention modules for better feature extraction. A lightweight module is designed for feature fusion to improve object detection across various scales.

Implementation Details

Training involves setting epochs to 300 with a batch size of 16, applying label smoothing to achieve optimal weights. The improved SRGAN algorithm features a generator and discriminator with multiple layers for processing image features efficiently. The model's feature extraction part includes attention modules that enhance detection accuracy. After training, the model is tested on a separate dataset to ensure its effectiveness in real-world applications.