[YOLOv5] Detailed explanation of each module of Backbone, Neck, and Head

Overview of YOLOv5 algorithm

Yolov5 is a target detection algorithm that uses an Anchor-based detection method and belongs to a single-stage target detection method. Compared with Yolov4, Yolov5 has faster speed and higher accuracy, and is currently one of the industry-leading target detection algorithms.

Basic principle of YOLOv5 algorithm

Yolov5 is based on the one-stage method in the target detection algorithm. Its main idea is to divide the entire image into several grids, and each grid predicts the type and position information of the object in the grid, and then according to the predicted frame and the real The IoU value between the boxes is used to filter the target box, and finally output the category and location information of the predicted box.

features

Yolov5 has the following characteristics:

• Efficiency: Compared with other target detection algorithms, Yolov5 is faster on the premise of ensuring high precision, especially in the GPU environment to achieve real-time detection.

• High accuracy: By using mechanisms such as multi-scale prediction and CIoU loss, Yolov5 can improve the accuracy of target detection.

• Ease of use: Yolov5 is open source and easy to use. It provides PyTorch version and ONNX version, which can run on different hardware.

Yolov5 can be applied to target detection tasks in various practical scenarios, such as object detection, face detection, traffic sign detection, animal detection, etc.

YOLOv5 model structure

There are five versions of yolov5: yolov5s, yolov5m, yolov5l, yolov5x and yolov5nano. Among them, yolov5s is the smallest version, and yolov5x is the largest version. The difference between them lies in the depth, width and parameter amount of the network.

The following uses yolov5s as a template to explain yolov5 in detail. It has higher precision and faster detection speed, and
at the same time has fewer parameters.
The YOLOv5s model is mainly composed of Backbone, Neck and Head. The network model is shown in the figure below. Among them:
Backbone is mainly responsible for feature extraction of the input image.
Neck is responsible for multi-scale feature fusion of feature maps and passing these features to the prediction layer.
Head does the final regression prediction.

Backbone backbone network

The backbone network refers to the network used to extract image features . Its main function is to convert the original input image into a multi-layer feature map for subsequent target detection tasks. In Yolov5, the CSPDarknet53 or ResNet backbone network is used. These two networks are relatively lightweight and can reduce the amount of calculation and memory usage as much as possible while ensuring high detection accuracy.
The main structures in Backbone are Conv module, C3 module, and SPPF module.

Conv module

The Conv module is a basic module commonly used in convolutional neural networks, which mainly consists of convolutional layers , BN layers and activation functions . These components are analyzed in detail below.

• The convolution layer is one of the most basic layers in the convolutional neural network, which is used to extract the local spatial information in the input features. The convolution operation can be regarded as a sliding window, the window slides on the input feature, and the feature value in the window is convolved with the convolution kernel to obtain the output feature. A convolutional layer usually consists of multiple convolution kernels, each corresponding to an output channel. Hyperparameters such as the size of the convolution kernel, step size, and padding method determine the output size and receptive field size of the convolution layer. In convolutional neural networks, convolutional layers are often used to build feature extractors.
• The BN layer is a normalization layer added after the convolutional layer to normalize the distribution of feature values ​​in the neural network. It can speed up the training process, improve model generalization ability, and reduce the model's dependence on initialization. The input of the BN layer is a batch feature map, which calculates the mean and variance of the features on each channel, and standardizes the features on each channel. The normalized features are restored by a learnable affine transformation (stretch and offset) to obtain the output of the BN layer.
• The activation function is a nonlinear function used to introduce nonlinear transformation capabilities to the neural network. Commonly used activation functions include sigmoid, ReLU, LeakyReLU, ELU, etc. They have different output behaviors in different ranges of input values, which can better adapt to different types of data distributions.

In summary, the Conv module is a commonly used basic module in convolutional neural networks. It extracts local spatial information through convolution operations, normalizes the distribution of eigenvalues ​​through the BN layer, and finally introduces nonlinear transformation capabilities through activation functions, so as to achieve Transformation and extraction of input features.

C3 module

The C3 module is an important part of the YOLOv5 network. Its main function is to increase the depth and receptive field of the network and improve the ability of feature extraction.

The C3 module is composed of three Conv blocks, where the first Conv block has a stride of 2, which can halve the size of the feature map, and the second and third Conv blocks have a stride of 1. The Conv block in the C3 module uses a 3x3 convolution kernel. Between each Conv block, BN layers and LeakyReLU activation functions are also added to improve the stability and generalization performance of the model.

The first Conv block in the C3 module has a stride of 2, and the two Convs in the red box form a Bottleneck, which means it will halve the size of the feature map. The purpose of this is to increase the receptive field of the network while reducing the amount of calculation. By halving the size of the feature map, the network can pay more attention to the global information of the object, thereby improving the effect of feature extraction.

The second and third Conv blocks in the C3 module have a stride of 1, which means they do not change the dimension of the feature map. The purpose of this is to maintain the spatial resolution of the feature map, so as to better preserve the local information of the object. At the same time, the main function of these two Conv blocks is to further extract features and increase the depth and receptive field of the network.

In general, the C3 module improves the ability of feature extraction by increasing the depth and receptive field of the network. This is very important for computer vision tasks such as object detection, because these tasks require accurate recognition and localization of objects, and accurate recognition and localization require good feature extraction capabilities.

SPP

The SPP module is a pooling module, usually used in convolutional neural networks, designed to achieve spatial invariance and position invariance of input data, so as to improve the recognition ability of neural networks. The main idea is to apply receptive fields of different sizes to the same image, so that feature information at different scales can be captured. In the SPP module, pooling operations of different sizes are first performed on the input feature maps to obtain a set of feature maps of different sizes. These feature maps are then concatenated and subjected to dimensionality reduction through a fully connected layer to finally obtain a fixed-size feature vector.

The SPP module usually consists of three steps:

• Pooling: The input feature maps are subjected to pooling operations of different sizes to obtain a set of feature maps of different sizes.
• Concatenation: Concatenate feature maps of different sizes together.
• Full connection: The dimensionality of the connected feature vector is reduced through the full connection layer to obtain a fixed-size feature vector.

Neck Feature Pyramid

Since the size and location of objects in an image are uncertain, a mechanism is needed to handle objects of different scales and sizes. Feature pyramid is a technique for dealing with multi-scale object detection, which can be achieved by adding feature layers of different scales on the backbone network. In Yolov5, the FPN (Feature Pyramid Network) feature pyramid structure is adopted, and the feature maps of different levels are fused together through upsampling and downsampling operations to generate a multi-scale feature pyramid. The top-down part mainly achieves the fusion of features from different levels by upsampling and fusion with coarser-grained feature maps, while the bottom-up part fuses feature maps from different levels by using a convolutional layer.

In the target detection algorithm, the Neck module is usually used to combine feature maps of different levels to generate feature maps with multi-scale information to improve the accuracy of target detection. In YOLOv5, a feature fusion module called PANet is used as the Neck module.

Specifically, the top-down part is to achieve the fusion of different levels of features by upsampling and fusion with coarser-grained feature maps, which are mainly divided into the following steps:

1.对最后一层特征图进行上采样，得到更精细的特征图；
2.将上采样后的特征图与上一层特征图进行融合，得到更丰富的特征表达；
3.重复以上两个步骤，直到达到最高层。

The bottom-up part mainly uses a convolutional layer to fuse feature maps from different levels, which is mainly divided into the following steps:

1.对最底层特征图进行卷积，得到更丰富的特征表达；
2.将卷积后的特征图与上一层特征图进行融合，得到更丰富的特征表达；
3.重复以上两个步骤，直到达到最高层。

Finally, the feature maps of the top-down part and the bottom-up part are fused to obtain the final feature map for object detection.

Head target detection head

The target detection head is the part used to perform target detection on the feature pyramid, which includes some convolutional layers, pooling layers, and fully connected layers. In the YOLOv5 model, the detection head module is mainly responsible for multi-scale target detection on the feature map extracted by the backbone network. This module mainly includes three parts. In addition, Yolov5 also uses some techniques to further improve the detection accuracy, such as GIoU loss, Mish activation function and multi-scale training.

• Anchors: Used to define target boxes of different sizes and aspect ratios, usually obtained by clustering the target boxes of the training set using K-means clustering, which can be calculated before model training and stored in the model for prediction Generate detection boxes.
• Classification: It is used to classify each detection frame to determine whether it is a target object. Usually, the feature is classified in the form of a fully connected layer plus a Softmax function.
• Regression: It is used to regress each detection frame to obtain its position and size, usually in the form of a fully connected layer to regress the features.

The detection head module of YOLOv5 adopts a multi-level feature fusion method. First, the feature map output by the backbone network is passed through a Conv module to reduce the number of channels and scale the feature map, and then fuse the feature maps of different levels to get more Rich feature information to improve detection performance.

Summary of YOLOv5

Yolov5 is a deep learning algorithm in the field of target detection. It is an improved version of Yolov4, which has achieved great improvements in speed and accuracy. The overall architecture of Yolov5 consists of backbone network, FPN, Neck, Head and other modules.
The backbone network part uses CSPDarknet53. By using the residual structure and feature reuse mechanism, the feature extraction ability of the model can be effectively improved.
The FPN part uses a feature pyramid based on Gaussian weighting, which can solve the problem of multi-scale target detection.

The Neck part adopts a structure combining SPP and PAN, which can improve the performance of the model while maintaining high efficiency.
The Head part uses the YOLOv5 head structure, which can output the prediction results of the network.

In general, the design of Yolov5 on each module fully considers the balance of speed and accuracy, making it perform well in target detection tasks.