For automatic Optical Inspection (AOI) machines that had Recognition been brought to printed Circuit Board market extra than five years in the past, illumination technique and mild gadgets are old. photographs captured by means of antique AOI machines aren’t easy to be identified via ordinary optical person popularity (OCR) algorithms, in particular for dark silk. the way to effectively growth silk reputation accuracy is quintessential for enhancing typical manufacturing performance in SMT plant. This paper makes use of quality tuned man or woman location cognizance for text Detection (CRAFT) method to construct model for dark silk recognition. CRAFT model consists of a shape similar to U-net, followed by way of VGG based totally convolutional neural network. non-stop -dimensional Gaussian distribution become used for the annotation of picture segmentation. CRAFT model is ideal at spotting distinct sorts of printed characters with excessive accuracy and transferability. consequences display that with the assist of CRAFT model, accuracy for good enough board is 95% (blunders rate is five%), and accuracy for NG board is 100% (omission charge is 0%).
Silk popularity, being one of the most critical parts of AOI inspection, plays a crucial position in monitoring produced PCB board. but, due to uneven heating in the reflow process, outdated illumination techniques, Recognition and insufficient lighting fixtures technology, silk popularity is particularly hard for most PCB plant life. mistakes charge is cease high that lots of correct forums are classified into awful boards, which results in hard human re-identity after AOI device.
the important thing technologies that the undertaking Recognition wishes to solve include a way to establish a wide and preferred model on restricted business defect samples; the way to imitate the way of taking into consideration the human brain to appropriately model commercial inspection gadgets which include silk display printing; the way to decrease both omission price and mistakes price in the identical time. In view of the issue of confined business defect samples, the era of “small samples to make large information” advanced by our organisation at the start of this 12 months is to mix deep antagonistic neural network with traditional computer imaginative and prescient technology to generate “limitless” facts with “restricted” facts. In view of the issue of imitating human mind wondering in modeling, we invite expert inspectors to label the statistics on the spot, mark the wrong factors for each defective sample that flows via, and assemble a convolutional neural community network model, learning and verifying pix, and performing reinforcement getting to know on misjudged pix to obtain accurate modeling. with a purpose Recognition to prevent defective merchandise from flowing out and enhance the skip charge of correct merchandise, our team has set parameters inside the model adjustable area, by placing dynamic tuning threshold to dam defective merchandise and enhance the skip fee of suitable merchandise.