Improvement 20 of yolov5: introduction of new neural network operator into network

​ front said ： As the current advanced deep learning target detection algorithm YOLOv5, A large number of trick, But there is still room for improvement , For the detection difficulties in specific application scenarios , There are different ways to improve . Subsequent articles , Focus on YOLOv5 How to improve is introduced in detail , The purpose is to provide their own meager help and reference for those who need innovation in scientific research or friends who need to achieve better results in engineering projects .

You are welcome to pay attention to more procedural information and answer questions —— WeChat official account ： Artificial intelligence AI Algorithm engineer  

solve the problem ： For classical convolution conv, It has always been a mainstay in neural networks , It has two characteristics ：spatial-agnostic and channel-specific（ Space agnostic and channel specific ）. In space ： The former ensures the efficiency of convolution kernel by reusing convolution kernel between different positions , And pursue translational equivalence . In the channel domain ： A series of convolution kernels are responsible for collecting all kinds of information encoded in different channels , Meet the latter characteristic . since VGGNet Since its appearance , Modern neural networks limit the space span of convolution kernel to no more than 3*3 To satisfy the compactness of convolution kernel .conv shortcoming :1、 It deprives the convolution kernel of the ability to adapt to different visual patterns in different spatial positions .2、 Locality limits the receptive field of convolution , Challenge small targets or blurred images .3、 The redundancy problem in convolution filter is prominent , The flexibility of convolution kernel is affected . In order to overcome these limitations of these classical convolutions , So the author of this paper puts forward involution The concept of involution During the operation , Inner volumes are space specific and channel agnostic “spatial-specific and channel-agnostic”（ And convolution contrary ）, The inner convolution nucleus is different in space , But it's shared on the channel , The authors also share on channel dimensions involution Core to reduce core redundancy . Dynamic parameterization involution The core has extensive coverage in spatial dimension . Through reverse design , What this article puts forward involution It has the double advantages of convolution :1、 Context semantic information can be aggregated in a broader space , Thus, the difficulty of modeling remote interaction is overcome ;2、 The weight can be adaptively distributed in different positions , Thus, the visual elements with the most abundant information in the spatial domain are prioritized .

principle ：

Thesis link ：https://arxiv.org/abs/2103.06255
github Code link ：https://github.com/d-li14/involution

Compared with the above standard convolution or depth convolution , Convolution kernel H The purpose is to include transformations with inverse characteristics in space and channel domain , Hence the name .（ Give... In your paper involution The explanation should be more consistent with “ Inverse characteristic convolution ” This Chinese translation . There are also many small partners on the Internet “ Combined volume ” perhaps “ Internal volume ”.）
stay involution In the design of ,involution The nucleus is specifically located in the corresponding coordinates (i,j) The pixel Xi,j custom , But sharing on the channel ,G Calculate that each group shares the same Involution The number of nuclear groups . utilize Involution Check the input for multiplication and addition , obtain Involution Of out feature map. Unlike convolution kernels ,Involution nucleus H The shape of the image depends on the input feature map X The shape of the . The design idea is to generate a matrix conditional on the original input tensor Involution nucleus , Align the output core with the input core .

Fang Law ：

First step ： modify common.py, Definition Involution modular .

class Involution(nn.Module):

    def __init__(self, c1, c2, kernel_size, stride):
        super(Involution, self).__init__()
        self.kernel_size = kernel_size
        self.stride = stride
        self.c1 = c1
        reduction_ratio = 4
        self.group_channels = 16
        self.groups = self.c1 // self.group_channels
        self.conv1 = Conv(
            c1, c1 // reduction_ratio, 1)
        self.conv2 = Conv(
            c1 // reduction_ratio,
            kernel_size ** 2 * self.groups,
            1, 1)

        if stride > 1:
            self.avgpool = nn.AvgPool2d(stride, stride)
        self.unfold = nn.Unfold(kernel_size, 1, (kernel_size - 1) // 2, stride)

    def forward(self, x):
        weight = self.conv2(self.conv1(x if self.stride == 1 else self.avgpool(x)))
        b, c, h, w = weight.shape
        weight = weight.view(b, self.groups, self.kernel_size ** 2, h, w).unsqueeze(2)
        # out = _involution_cuda(x, weight, stride=self.stride, padding=(self.kernel_size-1)//2)
        # print("weight shape:",weight.shape)
        out = self.unfold(x).view(b, self.groups, self.group_channels, self.kernel_size ** 2, h, w)
        # print("new out:",(weight*out).shape)
        out = (weight * out).sum(dim=3).view(b, self.c1, h, w)

        return out

The second step ： take yolo.py Register in Involution modular .

The third step ： change yaml file .

junction fruit ： I have done a lot of experiments on multiple data sets , Different data sets have different improvement effects .

Let me know ： Continue to share next YOLOv5 improvement . Interested friends can pay attention to me , If you have a problem, you can Leave a message or chat privately I oh ,csdn If you don't reply in time, please add WeChat official account ： Artificial intelligence AI Algorithm engineer .  

PS：Involution The method is not only applicable to improvement YOLOv5, You can also improve others YOLO The Internet , such as YOLOv4、v3 etc. .

Last , I hope I can powder each other , Be a friend , Learn and communicate together .

​