With the dramatic increase in the amount of garbage worldwide, garbage classification and recycling have become a key part of environmental protection and resource recycling. The aim of this project is to develop a machine learning-based garbage recognition system for the automatic classification of daily garbage. We adopt deep learning methods, especially Convolution Neural Networks to recognize and classify various kinds of garbage. By collecting and processing a large amount of garbage image data, we trained a high-precision model and performed several rounds of optimization on it. Preliminary test results show that the system can effectively recognize and classify different kinds of garbage and provide a decision basis for its automatic sorting. The successful implementation of this system will not only improve the efficiency of garbage sorting but also encourage society to make more progress in garbage management and resource recovery.
Object detection is one of the cornerstone tasks in the field of computer vision, playing a pivotal role in many contemporary applications like autonomous driving, medical image analysis, and automated surveillance. With the growing global garbage, applying object detection techniques to automated garbage classification becomes increasingly imperative. While the YOLO series has excelled in various detection scenarios, the detection of small objects remains a vexing challenge. In addressing this, we present an augmented YOLO model tailored for enhanced small object detection. We aim to validate its efficacy through garbage detection tasks, hoping that it not only boosts garbage detection accuracy but also brings substantial value to other related domains. Moreover, advancements in tech-driven sorting, such as AI and robotics, offer prospects for smart city waste management, enhancing efficiency and reducing environmental footprints.
Input: An image containing various types of garbage, potentially with small objects.
Output: Detection labels for the garbage items in the image.
To tackle small object detection in garbage classification, we’ll test and compare three primary strategies:
-
SPD-Conv Module: Modifying convolution and pooling for better feature detail.
-
SAHI Framework: A flexible slicing-aided inference pipeline tailored for small object detection.
-
YOLOv8 Enhancements: Integrating extra prediction heads, Involution blocks, and CBAM mechanisms.
The object detection domain, particularly focusing on small objects, has been an area of active research. The YOLO series introduced fast and efficient detection models. However, specific challenges like detecting smaller objects led to advancements like the SPD-Conv module , the SAHI framework , and enhancements in YOLOv5 . Our work builds on these foundations but introduces unique modifications for garbage classification.
-
Assessing the model’s accuracy using the size, mAP, params(M), and FLOPs(B).
-
Displaying visual results to provide an intuitive grasp of the model’s performance in real-world scenarios, including successful and failed detections.