YOLOv5 Neural Network Small Object Detection via CIoU-NMS
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection and recognition methods based on neural networks face challenges in accurately detecting small objects and multiple objects closely arranged, leading to false negatives and positives, and struggle to meet real-time detection requirements.
Innovation Solution
A modified YOLOv5 network model is constructed by adding a detection layer, using a specific training process that includes setting thresholds for center-to-center distance and aspect ratio, and employing the CIoU-NMS method for non-maximum suppression, which improves detection accuracy and speed by considering overlapping areas, distances, and aspect ratios between predicted and ground truth boxes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If region proposal methods (SPP, R-CNN, Fast R-CNN) are used for object detection, then detection accuracy is improved, but detection speed deteriorates and real-time performance cannot be met
Solution Approach 1:
The patent segments the detection task into multiple stages by introducing intermediate detection layers between the backbone and output layers. This allows the model to process images in a hierarchical manner, extracting features at different scales and then performing detection, which balances accuracy and speed by avoiding the need to process all candidate regions at full resolution.
Solution Approach 2:
The patent adds detection layers at different dimensional levels (different depths in the network) to capture objects of varying sizes. By operating at multiple dimensional levels simultaneously, the model can maintain high accuracy for small objects while preserving the computational efficiency needed for real-time detection.
2Productivity
If end-to-end methods (SSD, RetinaNet, YOLOv3) are used for object detection, then detection speed is improved, but detection accuracy deteriorates
Solution Approach 1:
The patent divides the single-stage end-to-end detection into multiple stages by inserting intermediate detection layers. This segmentation allows the model to perform preliminary detections at coarser levels and refine them at finer levels, thereby improving overall accuracy while maintaining the fast inference speed of end-to-end methods.
Solution Approach 2:
The patent applies different detection strategies at different locations and depths within the network. Intermediate detection layers process different regions of the feature pyramid with different levels of detail, allowing the model to allocate computational resources according to local requirements and improve overall detection accuracy without sacrificing speed.
3Productivity
If YOLOv5 network model is used, then detection accuracy and speed are improved, but performance on small object extraction deteriorates
Solution Approach 1:
The patent adds detection layers at higher dimensional levels (deeper in the network) to capture small objects that were previously missed. By operating at multiple depth levels, the model can simultaneously detect both large and small objects, resolving the limitation of single-stage YOLOv5 in extracting small objects.
Solution Approach 2:
The intermediate detection layers perform preliminary detection of small objects at earlier stages before the main detection occurs. This preliminary action allows the model to identify and process small objects that would otherwise be lost in the downsampling process, improving small object detection accuracy while maintaining overall speed.
Data Source
AI summary
The present disclosure provides a method for object detection and recognition based on a neural network. The method includes: adding a detection layer following three detection layers of an existing YOLOv5 network model, to construct a new YOLOv5 network model; then, training the new YOLOv5 network model by considering an overlapping area between a predicted box and a ground truth box, a center-to-center distance between the two boxes, and an aspect ratio of the two boxes; and finally, inputting a to-be-detected image into the trained new YOLOv5 network model, outputting a predicted box of an object and probability values corresponding to a class to which the object belongs, and setting a class corresponding to a maximum probability value as a predicted class of the object in the to-be-detected image. This method can quickly and effectively detect multiple classes of objects. Especially, a detection effect for small objects is more ideal.