重参数化大核卷积的光学黑色素瘤图像检测算法

Optical melanoma image detection algorithm based on heavy parameterized large kernel convolution

摘要:

针对皮肤镜采集的光学黑色素瘤图像,由于其背景信息复杂,干扰信息过多,导致检测精度较低,出现误检、漏检等问题。因此提出一种重参数化大核卷积的光学黑色素瘤图像检测算法。首先,在主干部分设计一种融合大核卷积与C3的新模块C3_RepLK,增大模型的感受野,提取更多的有效信息。其次,引入感受野模块RFB,融合不同尺度的特征信息,减少错检问题。颈部网络中采用混合密集稀疏卷积GSConv和轻量化上采样算子CARAFE,使得网络能够捕捉到丰富的上下文信息,抑制漏检问题。最后,在算法中融入二阶通道注意力模块SOCA,加强特征之间的关联性,关注更有用的特征。实验表明,所提检测算法较原YOLOv5算法,所有类别平均精度从85.0%提升至89.4%,证明了所提出的算法对于检测黑色素瘤的有效性。

Abstract:

Due to the complex background information and excessive interference information, the detection accuracy of optical melanoma images collected by dermoscope is low, and there are some problems such as false detection and missing detection. Therefore, an optical melanoma image detection algorithm based on heavy parametric large nuclear convolution is proposed. Firstly, a new module C3_RepLK, which combines large nuclear convolution with C3, is designed in the backbone to increase the receptive field of the model and extract more effective information. Secondly, the receptive field module RFB is introduced to fuse the feature information of different scales to reduce the problem of error detection. Mixed dense and sparse convolutional GSConv and lightweight upsampling operator CARAFE are used in the neck network, which enables the network to capture rich context information and suppress the problem of missing detection. Finally, the second-order channel attention module SOCA is incorporated into the algorithm to strengthen the correlation between features and focus on more useful features. Experiments show that compared with the original YOLOv5 algorithm, the average accuracy of all categories is improved from 85.0% to 89.4%, which proves the effectiveness of the proposed algorithm in detecting melanoma.