[target detection] yolov5, the shoulder of target detection (detailed principle + Training Guide)

1.YOLO Input end

1.1 Mosaic Data to enhance

Yolov5 And... Are used at the input of Yolov4 Same Mosaic The way data is enhanced .

Mosaic The author of the data enhancement proposal is also from Yolov5 Team members , however , Random scaling 、 Random cutting 、 Random layout of the way to splice , The detection effect of small targets is still very good .

• 4 Picture mosaic
• Random scaling
• Random cutting
• Random arrangement

Algorithm advantages ：

• Rich data set
• Reduce GPU Calculation

1.2 Adaptive anchor frame calculation

stay Yolo In the algorithm, , For different data sets , There will be an anchor frame with initial set length and width .

In network training , The network outputs the prediction box based on the initial anchor box , And then with the real box groundtruth compare , Calculate the difference between the two , Then reverse update , Iterative network parameters .

• Step1： Read all pictures in the training set w、h And the detection frame w、h

• Step2： Correct the read coordinates to absolute coordinates

• Step3： Use Kmeans The algorithm clusters all the detection frames in the training set , obtain k individual anchors

• Step4： Through genetic algorithm to get anchors To mutate , If the effect of mutation is good, keep it , Otherwise, skip

• Step5： Will ultimately get the best anchors Return by area

1.3 Adaptive image scaling

In common target detection algorithms , Different pictures have different lengths and widths , Therefore, the common way is to uniformly scale the original image to a standard size , Then send it to the detection network .

letterbox Adaptive image scaling technology tries to maintain the aspect ratio , The missing parts shall be filled with gray edges to reach the fixed size .

2.YOLO Overall architecture

2.1 BackBone

Mainly for feature extraction , The object information in the image is extracted through convolution network , Used for later target detection .

2.1.1 Focus modular

Focus Layer principle and PassThrough Layers are very similar . It uses slicing operations to split high-resolution images into multiple low-resolution images / Characteristics of figure , That is, interlaced sampling + Splicing .

2.1.2 SPP modular

Space Pyramid pooling , It can convert any size feature map into a fixed size feature vector .

2.1.3 CSP_X modular

backbone It's a deeper network , Increasing the residual structure can increase the gradient value of back propagation between layers , Avoid the disappearance of gradients caused by deepening , Thus, finer grained features can be extracted without worrying about network degradation .

2.2 Neck

Mix and combine features , Enhance the robustness of the network , Strengthen the ability of object detection , And pass these characteristics to Head Layer to predict .

2.2.1 FPN

2.2.2 PAN

2.3 YOLO Output terminal

It mainly carries out the final prediction output .

2.3.1 Bounding Box Loss function

The consistency between the real detection box and the model prediction output box , For back propagation optimization model .

2.3.2 NMS Non maximum suppression

Judge whether the adjacent grids recognize the same object , Eliminate redundant detection boxes .