Jelly impurity detection based on scalable overlapping slide pooling SSD network

An impurity detection method based on SSD is proposed to detect the internal impurity of jelly in the production line. In the pre-processing stage, a sliding image block segmentation method is proposed to segment the whole jelly image into several segmentation block images, so as to avoid the low accuracy because the image is too large and the target is too small. Using the method of transfer learning, the features learned from the neural network on ImageNet are transferred to the jelly database to accelerate the convergence speed of the network and avoid over-fitting to a certain extent. A scalable overlapping slide pooling (SOSP) method is proposed to replace the fifth pooling to achieve more robust feature pooling. Finally, the detection results on all segmentation block images under one image are integrated to obtain the detection results of the whole image. Experimental results show that the proposed method is effective and feasible. The average accuracy for the multiple defects reaches 0.7271. Compared with other methods, our algorithm is more robust and can be applied to the jelly impurity detection production line.