To ensure the normal operation of the automatic knotting machine of yarn in the automatic bobbin changing system, we need to detect the yarn sucked by the pipe. Yarn is detected by image processing instead of sensors because it is thin with diverse types and colors. However, traditional image processing methods are too complex and inaccurate to identify yarn with various types, sizes, and colors. This study proposes a network of multi-scale depth separable convolution blocks modified based on Inception-Resnet-A block of Inception v4 to detect yarn in pipes. The conventional 3×3 convolution layers in the Inception-ResNet-A block is replaced with the depth separable convolution layers of the 3×3 convolution kernel, and some of the 1×1 convolution layers are removed for less parameters of convolution blocks and simpler calculation. In addition, ResNet is employed for channel fusion to prevent feature loss. According to the experimental results, this network model is remarkable in generalization and recognition.