Here I mainly want to record and summarize the principles, advantages and disadvantages of various models in the field of target detection that are usually intentionally or unintentionally checked, learned, and so on.
Many classic models have emerged for target detection tasks. The following is a summary record of a model.
(1) R-CNN (Region-based Convolutional Neural Networks):
algorithm principle: R-CNN first generates candidate regions, then performs convolutional neural network feature extraction on each candidate region, and uses support vector machine (SVM) for classification . Finally, the detection results are precisely localized using bounding box regression.
Advantages: high accuracy, able to detect small targets.
Disadvantages: The speed is relatively slow, the detection process is divided into multiple stages, and the computational complexity is high.
(2) Fast R-CNN:
Algorithm principle: Fast R-CNN uses a fully convolutional network to extract image features and takes the entire image as input. At the same time, a region pooling operation is introduced to align candidate regions of different sizes onto a fixed-size feature map. Finally, classification and bounding box regression are performed through a fully connected layer.
Advantages: Compared with R-CNN, it is faster and can detect targets in real time.
Disadvantages: The training process is relatively complicated and requires additional region pooling operations.
(3) Faster R-CNN:
Algorithm principle: Faster R-CNN proposes a Region Proposal Network (RPN) for generating candidate regions. The RPN and the entire detection network share the convolutional feature extraction part, and generate region proposals and object detection through anchor boxes and classification/regression heads.
Advantages: faster speed and higher accuracy.
Disadvantages: The training process is relatively complicated and requires an additional region generation network.
(4) Mask R-CNN:
Algorithm principle: Mask R-CNN introduces a fully convolutional network based on Faster R-CNN to achieve accurate instance segmentation. On top of object detection, a per-pixel semantic segmentation mask is also predicted via an additional output.
Advantages: Ability to simultaneously perform target detection and semantic segmentation tasks.
Disadvantages: The amount of calculation is large, and the requirements for hardware resources are high.
(5) YOLO (You Only Look Once):
Algorithm principle: YOLO transforms the target detection problem into a regression problem, and performs dense prediction directly on the image. It divides an image into grid cells and within each cell predicts multiple bounding boxes and class probabilities.
Advantages: very fast, suitable for real-time applications.
Disadvantages: The positioning accuracy is relatively low, and the detection effect on small targets is not ideal.
In addition, YOLO, as a very famous model in the field of target detection tasks, has evolved from the original v1 model to the latest v8 version so far. Next, the models will be introduced in turn.
YOLOv1:
Algorithm principle: YOLOv1 adopts a fully convolutional network structure, which simultaneously performs target detection and classification in a single forward pass, transforming the target detection problem into a regression problem. It works by dividing the input image into smaller grids and predicting bounding box coordinates and class probabilities at each grid.
Advantages: YOLOv1 has a faster speed and can perform target detection in real time. It can detect multiple targets, and it does a better job of detecting small targets.
Disadvantages: YOLOv1 is relatively poor in positioning accuracy, and is prone to errors in the position and size of the bounding box. In addition, since a single-scale feature map is used for prediction, the object detection effect for different scales is weak.
YOLOv2 (YOLO9000):
Algorithm principle: YOLOv2 has been improved on the basis of YOLOv1, introducing Anchor Boxes and multi-scale prediction. It uses Anchor Boxes to provide more accurate bounding box predictions, and handles objects of different scales by making predictions on different levels of feature maps.
Advantages: YOLOv2 has higher positioning accuracy and detection performance than YOLOv1. It can handle targets of different scales, and has good adaptability and versatility.
Disadvantages: Compared with some of the latest models, YOLOv2 still has room for improvement in performance on small target detection and high-resolution images.
YOLOv3:
Algorithm principle: YOLOv3 has been further improved on the basis of YOLOv2, introducing FPN structure and multi-scale prediction. It uses the FPN structure to fuse feature maps at different levels to capture semantic information at different scales, and handle objects at different scales by predicting bounding boxes of different sizes.
Advantages: YOLOv3 has significantly improved accuracy and speed, and has better detection capabilities for targets of different scales and sizes. It also adopts more training tricks and data augmentation strategies to improve the robustness of the model.
Disadvantages: Compared with some latest models, such as YOLOv4 and YOLOv5, YOLOv3 still has room for improvement in the performance of small target detection and specific scenarios.
YOLOv4:
Algorithm principle: YOLOv4 is an important version in the YOLO series, which has been improved in terms of network structure, loss function, and data enhancement. YOLOv4 introduced CSPDarknet53 as the backbone network, using more convolutional layers and residual connections to extract features. In addition, it also uses IoU loss function, Mosaic data enhancement and other techniques to improve the detection performance.
Advantages:
Advantages: YOLOv4 has achieved a significant improvement in object detection performance. It has higher accuracy and better localization ability, especially in small object detection and dense object detection. At the same time, YOLOv4 also introduces some new technologies, such as CIOU loss, GIoU loss and SAM module, etc., which further improve the accuracy and robustness of the model.
Disadvantages: YOLOv4 is more complex than the previous version, and requires higher computing resources for training and reasoning. This makes it highly demanding on hardware devices, and may not be suitable for resource-constrained or high-real-time requirements
YOLOv5:
Algorithm principle: YOLOv5 is the latest version of the YOLO series, which improves by introducing the ideas of EfficientDet and PANet. YOLOv5 uses a lightweight network structure, and uses multi-level feature fusion and cross-stage feature pyramid methods to improve the performance of target detection.
Advantages: YOLOv5 inherits the real-time and high efficiency of the YOLO series, and at the same time has a greater improvement in accuracy. It has good performance on small and dense object detection, and can achieve efficient object detection on resource-constrained devices.
Disadvantages: Compared with some of the latest target detection models, such as DETR and EfficientDet, YOLOv5 still has room for improvement in terms of robustness and inference speed in specific scenarios. Also, YOLOv5 is developed by an unofficial team of developers.
(6) SSD (Single Shot MultiBox Detector):
Algorithm principle: SSD also converts the target detection problem into a regression problem, by predicting multiple bounding boxes and category probabilities on feature maps of different scales. It uses multiple layers of features to detect objects at different scales.
Advantages: faster speed, able to detect targets of different scales.
Disadvantages: Relative to other models, the accuracy is slightly lower.
(7) RetinaNet:
Algorithm principle: RetinaNet proposes an effective method to solve the problem of category imbalance in target detection. It uses a special loss function to balance the weight between positive and negative samples, making the model pay more attention to samples that are difficult to classify. In addition, RetinaNet uses a multi-scale feature pyramid network (FPN) to detect objects at different levels.
Advantages: It has a good effect on category imbalanced data sets and can detect small targets; using multi-scale feature pyramids can adapt to targets of different scales.
Disadvantages: Compared with some simplified models, the computational complexity is higher.
(8) EfficientDet:
Algorithm principle: EfficientDet realizes target detection by combining the improved EfficientNet feature extractor and BiFPN (Bi-directional Feature Pyramid Network) feature fusion module. It also considers the balance of detection speed and accuracy.
Advantages: A good balance has been achieved in terms of speed and accuracy; through the EfficientNet feature extractor, it has a high feature representation ability.
Cons: Relatively new model, may require more tuning and adaptation in some cases.
(9) Cascade R-CNN:
Algorithm principle: Cascade R-CNN uses multiple R-CNN modules in cascade to improve the accuracy of target detection. In each cascading stage, the model screens out more accurate candidate boxes and trains again.
Advantages: Compared with other models, it performs better in terms of accuracy and can detect small targets; the cascade structure helps to further improve performance.
Disadvantages: High model complexity, long training and inference time.
(10) FCOS (Fully Convolutional One-Stage Object Detection)
algorithm principle: FCOS is a fully convolutional single-stage object detection algorithm. It achieves object detection by predicting object classification and bounding box information for each pixel on the entire feature map. FCOS adopts an anchor-free design, transforms detection problems into pixel-level classification and regression tasks, and performs multi-scale target detection in a layer-by-layer top-down manner.
Advantages:
There is no need to manually define the anchor, which reduces the complexity of the prior box setting.
The fully convolutional architecture allows the input of the network to be an image of any size.
It has a better effect on small target detection.
It has high positioning accuracy on large objects.
Disadvantages:
large amount of calculation, relatively slow in terms of speed.
The detection effect for dense objects is relatively poor.
(11) Principle of CornerNet
algorithm: CornerNet is a corner-based target detection algorithm. It uses two parallel networks to predict the corners and center points of the target respectively, and then obtains the final detection result by matching the corner points and center points. CornerNet uses a reverse stacked Hourglass network to extract features and predict corners and centers.
Advantages:
In object detection, corner information provides richer geometric information, which can effectively solve problems such as occlusion and rotation.
It is better for small target detection.
Faster in terms of speed.
Disadvantages:
Relatively poor positioning accuracy on large objects.
The detection effect on dense objects is not good.
(12) Principle of CenterNet
algorithm: CenterNet is a target detection algorithm based on center points. It achieves object detection by predicting the center point, width and height of the object. CenterNet uses a simple convolutional network to extract features, and predicts the center point and width and height through an additional regression head.
Advantages:
Simple and lightweight structure, with great advantages in speed and memory consumption.
The detection effect is better for small targets and dense targets.
It has better robustness in complex scenes.
Disadvantages:
The detection effect for large objects is relatively poor.
There may be a certain loss in positioning accuracy.
(13) DETR (Detection Transformer)
algorithm principle: DETR is a Transformer-based target detection algorithm. It transforms the object detection problem into a sequence-to-sequence problem, generating object categories and bounding boxes through an encoder-decoder structure. DETR uses an attention mechanism to build global context information, and introduces object location encoding and self-attention mechanisms to achieve object detection.
Advantages:
The category and bounding box of the target can be directly generated without using prior information such as anchor.
It has a better effect when dealing with occlusion and dense targets.
It has high positioning accuracy.
Disadvantages:
The detection effect for small targets is relatively poor.
The training process requires a large number of samples and computing resources.
(14) FPN (Feature Pyramid Network)
algorithm principle: FPN is a feature pyramid network used to solve multi-scale problems in target detection. It provides multi-scale feature representation, from low-level to high-level, by performing information fusion and upsampling operations on feature maps at different levels. FPN combines high-resolution low-level features with semantic-rich high-level features through top-down and lateral connections, so as to preserve detailed information and capture more abstract semantic information.
Advantages:
It provides multi-scale feature representation and adapts to the detection requirements of targets of different scales.
It has better performance in object localization and multi-scale detection.
It can be used as the backbone network structure of other object detection models.
Disadvantages:
FPN mainly focuses on feature representation, and its reasoning ability at the object level is relatively weak.
Does not directly address issues such as small object detection.
(15) Principle of MobileNet-SSD algorithm:
MobileNet-SSD is a lightweight target detection algorithm that uses MobileNet as a feature extractor and combines it with Single Shot MultiBox Detector (SSD) for target detection. MobileNet-SSD achieves target detection by predicting bounding boxes and category scores at different scales through convolutional layers.
Advantages:
Lightweight structure, suitable for resource-constrained devices and real-time detection applications.
Faster inference speed.
It has good detection ability in general scenarios.
Disadvantages:
Compared with some complex models, certain detection accuracy may be sacrificed.
It is not suitable for special scenes that require high precision.