In order to enable the quayside crane surface corrosion detection task to be deployed to embedded devices and mobile devices to obtain faster inference speed, an improved lightweight object detection network MobileNetV2SSDLiteV1/V2 is proposed without sacrificing accuracy. The improved network uses feature maps of 5 convolutional layers as the detector input, and uses 3×3 depthwise convolution to predict classification and location scores. In order to reduce the number of parameters of the backbone network, the original 17 inverted linear bottleneck block structure is designed into 14, and the image with a resolution of 256 pixel × 256 pixel is used as the network input to change the coefficients of the original default box. The number of is reduced by 82.51%, and then all convolutions are normalized and the network is trained from scratch. The above improvements can make the network parameters become 0.96×10 ⁶, which is reduced to 1/4 of the original. The number of floating-point operations of the network is 0.12×10 ⁹, which is 81.25% less than the original, the mAP value is as high as 77.40%, and the inference speed reaches 45 frame/s.