【Deep Learning OCR Series · 7】CTC Loss Function and Training Techniques

## Okwu Mmalite Connectionist Temporal Classification (CTC) bụ ọganihu dị mkpa na usoro mmụta miri emi, ọkachasị n'ọhịa nke OCR. CTC na-edozi nsogbu bụ isi nke nkwekọrịta n'etiti ogologo usoro ntinye na usoro mmepụta, na-eme ka mmụta usoro njedebe na njedebe. Isiokwu a ga-abanye n'ime ụkpụrụ mgbakọ na mwepụ, mmejuputa algorithm, na usoro njikarịcha ọzụzụ nke CTC. ## CTC Basic Concepts ### Usoro nhazi nsogbu N'ihe banyere ọrụ OCR, anyị na-eche ihe ịma aka ndị a: ** Ogologo nkwekọrịta **: Ogologo nke usoro ihe oyiyi ntinye dị iche na ogologo ederede ederede. Dịka ọmụmaatụ, okwu nwere mkpụrụedemede 3 nwere ike kwekọọ na usoro atụmatụ nke usoro oge 100. * Ọnọdụ a na-ejighị n'aka: A maghị kpọmkwem ọnọdụ nke agwa ọ bụla na onyinyo ahụ. Usoro ọdịnala chọrọ nkewa agwa ziri ezi, nke siri ike na ngwa bara uru. ** Ihe isi ike na Character Segmentation **: Ederede ederede na-aga n'ihu, ederede ejiri aka dee, ma ọ bụ mkpụrụedemede nka na-agbasi mbọ ike kewaa n'ụzọ ziri ezi n'ime mkpụrụedemede ọ bụla. ### Ngwọta CTC CTC na-edozi nsogbu nhazi usoro n'ụzọ ọhụrụ ndị a: Na-ewebata Blank Markers: Jiri akara ngosi pụrụ iche na-acha ọcha iji jikwaa nkwekọrịta. Mkpado oghere adabaghị na mkpụrụedemede mmepụta ọ bụla ma jiri ya kewaa mkpụrụedemede oyiri site na usoro jupụta. Pathzọ puru omume: Gbakọọ ihe puru omume nke niile kwere omume nkwekọrịta ụzọ. Ụzọ ọ bụla na-anọchite anya usoro mmekọrịta mmadụ na ibe ya. ** Dynamic Planning **: Gbakọọ nke ọma ụzọ puru omume site na iji algọridim na-aga n'ihu-azụ, na-ezere ịgụta ụzọ niile enwere ike. ## CTC Mathematics Principles ### Nkọwa ndị bụ isi Nyere usoro ntinye X = (x₁, x₂, ..., xt) na usoro lekwasịrị anya Y = (y₁, y₂, ..., yu), ebe T ≥ U. Mkpado setịpụrụ: L = {1, 2, ..., K}, nwere K agwa edemede. ** Extended Tag Collection **: L_ext = L ∪ {blank}, nwere mkpado oghere. ** Ụzọ nhazi **: Usoro nke ogologo T π = (π₁, π₂, ..., πt), ebe πt ∈ L_ext. ### Map nke ụzọ na mkpado CTC na-akọwa ọrụ eserese B nke na-agbanwe ụzọ nhazi n'ime usoro akara mmepụta: 1. Wepụ ihe niile na-acha ọcha 2. Jikọta ihe odide oyiri na-esote ** Ihe Nlereanya Maapụ **: - π = (a, a, blank, b, blank, b, b) → B (π) = (a, b, b) - π = (oghere, c, c, a, oghere, t) → B (π) = (c, a, t) ### CTC ọnwụ ọrụ A na-akọwa ọrụ ọnwụ CTC dị ka logarithm na-adịghị mma nke nchikota nke ụzọ niile nke puru omume na-emepụta usoro Y: L_CTC = -log P(Y| X) = -log Σ_{π∈B⁻¹(Y)} P(π| X) ebe B⁻¹ (Y) bụ setịpụrụ ụzọ niile edepụtara na Y. Patha puru omume: Na-eche na amụma nke oge ọ bụla na-anọghị n'onwe ya, ụzọ nke puru omume bụ: P(π| X) = ∏t yt^{πt} ebe yt^{πt} bụ ihe puru omume nke oge nzọụkwụ t na-ebu amụma akara πt. ## Aga n'ihu-azụ algọridim ### Aga n'ihu algọridim Algọridim na-aga n'ihu na-agbakọ ụzọ nke puru omume site na mmalite nke usoro ahụ ruo n'ọnọdụ dị ugbu a. ** Extended Label Sequence **: Iji kwado ngụkọta, gbasaa usoro Y ruo Y_ext, na-etinye mkpado efu tupu na mgbe agwa ọ bụla. ** Mbido **: - α₁(1) = y₁^{blank} (ọnọdụ mbụ bụ oghere) - α₁(2) = y₁^{y₁} (ọnọdụ mbụ bụ agwa mbụ) - α₁(s) = 0 maka ebe ndị ọzọ ** Recursive Formula **: Maka t > 1 na ọnọdụ s: - Ọ bụrụ na Y_ext [s] na-acha ọcha ma ọ bụ otu ihe ahụ dị ka agwa gara aga: α_t(s) = (α_{t-1}(s) + α_{t-1}(s-1)) × y_t^{Y_ext[s]} - Ma ọ bụghị ya: α_t(s) = (α_{t-1}(s) + α_{t-1}(s-1) + α_{t-1}(s-2)) × y_t^{Y_ext[s]} ### Backward Algorithm Algọridim azụ na-agbakọ ụzọ nke puru omume site na ọnọdụ dị ugbu a ruo na njedebe nke usoro ahụ. ** Mbido **: - β_T(| Y_ext|) = 1 - β_T(| Y_ext|-1) = 1 (ma ọ bụrụ na mkpado ikpeazụ abụghị oghere) - β_T(s) = 0 maka ebe ndị ọzọ ** Recursive Formula **: N'ihi na t < T na ọnọdụ S: - Ọ bụrụ na Y_ext [s + 1] bụ ihe efu ma ọ bụ otu ihe ahụ dị ka agwa dị ugbu a: β_t(s) = (β_{t+1}(s) + β_{t+1}(s+1)) × y_{t+1}^{Y_ext[s+1]} - Ma ọ bụghị ya: β_t(s) = (β_{t+1}(s) + β_{t+1}(s+1) + β_{t+1}(s+2)) × y_{t+1}^{Y_ext[s+1]} ### Ngụkọta Gradient Ngụkọta nke puru omume: P (Y | X) = α_T(| Y_ext|) + α_T(| Y_ext|-1) ** Gradient nke Label Probability**: ∂(-ln P(Y| X))/∂y_k^t = -1/P(Y| X) × Σ_{s:Y_ext[s]=k} (α_t(s) × β_t(s))/y_k^t ## CTC decoding atụmatụ ### Anyaukwu decoding Anyaukwu na-ekpughe akara ahụ nwere ohere kachasị elu n'oge ọ bụla: π_t = argmax_k y_t^k Mgbe nke ahụ gasịrị, pịa bọtịnụ B iji nweta usoro ikpeazụ. * Nkowasi: Jikwaa gị ngwa ngwa na ngwa * Ngwọta zuru ụwa ọnụ nwere ike ọ gaghị ekwe omume. ### Ngwugwu ọchụchọ decoding Beam search na-ejigide ọtụtụ ụzọ ndị na-aga ime, na-agbasawanye ụzọ kachasị ekwe nkwa n'oge ọ bụla. ** Nzọụkwụ algorithm **: 1. Mbido: Nchịkọta nwa akwukwo nwere ụzọ efu 2. Maka oge ọ bụla nzọụkwụ: - Gbasaa ụzọ niile na-aga ime Debe ụzọ K-ụzọ na ohere kachasị elu. 3. Weghachite ụzọ zuru oke na nke puru omume kachasị elu ** Parameter Tuning **: - Beam obosara K: Itule mgbagwoju anya mgbakọ na kọmitii na decoding àgwà - Ntaramahụhụ ogologo: Zere ịkwado usoro dị mkpirikpi ### Prefix bundle search Prefix ngwugwu search na-atụle prefix puru omume nke a ụzọ iji zere abụọ-agụta ụzọ na otu prefix. ** Isi echiche **: Jikọta ụzọ na otu prefix, ma na-edebe usoro ndọtị kachasị mma. ## Ọzụzụ ọzụzụ na njikarịcha ### Data preprocessing ** Usoro Ogologo Nhazi **: - Dynamic batching: Grouping usoro nke yiri ogologo - Fill Strategy: Jupụta obere usoro na pụrụ iche markers - Truncation Strategy: Ezi uche truncate gabigara ókè ogologo usoro ** Label Preprocessing **: - Character Set Standardization: Uniform character encoding and capitalization - Njikwa agwa pụrụ iche: Na-ejikwa akara akara na oghere - Vocabulary Building: Build a zuru ezu glossary nke odide ### Ọzụzụ Ọzụzụ ** Ọmụmụ mmụta **: Malite ọzụzụ na ihe atụ dị mfe ma jiri nwayọọ nwayọọ mụbaa ihe isi ike: - Short to Long Sequences - Clear image to blurry image - Regular fonts to handwritten fonts ** Nkwalite data **: - Geometry mgbanwe: bugharịa, ọnụ ọgụgụ, bee - Mkpọtụ mkpọtụ: Mkpọtụ Gaussian, nnu na ose mkpọtụ - Mgbanwe ọkụ: nchapụta, mgbanwe ọdịiche ** Usoro nhazi **: - Dropout: Gbochie overfitting - Ibu mmebi: L2 regularization - Label Smoothing: Na-ebelata ntụkwasị obi gabigara ókè ### Nhazi Hyperparameter ** Usoro mmụta **: - Usoro okpomọkụ: Oge ole na ole mbụ na-eji obere mmụta - Cosine annealing: Ọnụego mmụta na-emebi emebi dịka ọrụ cosine si dị. - Adaptive Tuning: Na-agbanwe dabere na nkwenye setịpụrụ arụmọrụ ** ogbe size nhọrọ **: - Ebe nchekwa nchekwa: Tụlee ikike ebe nchekwa GPU - Gradient Stability: Na-enye gradient kwụsiri ike maka nnukwu ogbe - Convergence Speed: Itule ọzụzụ ọsọ na nkwụsi ike ## Echiche Ngwa Bara Uru ### Njikarịcha Mgbakọ ** Njikarịcha ebe nchekwa **: - Gradient checkpoints: Na-ebelata akara ukwu ebe nchekwa nke mgbasa ozi n'ihu - Mixed-nkenke ọzụzụ: Belata ebe nchekwa chọrọ na FP16 - Dynamic graph njikarịcha: Optimizes ebe nchekwa nkesa maka gbakọọ eserese ** Njikarịcha ọsọ **: - Parallel Computing: Na-eji GPU parallel nhazi ike - Algorithm njikarịcha: Emejuputa atumatu site na iji oru oma n'ihu-na-azụ algọridim - ogbe njikarịcha: Tọọ ogbe nha n'ụzọ kwesịrị ekwesị ### Ọnụ ọgụgụ kwụsie ike ** Ngụkọta oge puru omume **: - Log-ohere ngụkọta: Zere uru overflow mere site puru omume multiplication - Numeric clipping: Na-egbochi nso nke puru omume ụkpụrụ - Normalization Techniques: Hụ na izi ezi nke puru omume nkesa ** Gradient kwụsie ike **: - Gradient Cropping: Na-egbochi mgbawa gradient - Weight Initialization: Jiri usoro mmalite kwesịrị ekwesị - Batch normalization: stabilizes ọzụzụ usoro ## Nyocha arụmọrụ ### Nyochaa metrik ** Character-larịị ziri ezi **: Accuracy_char = Ọnụ ọgụgụ nke ihe odide a ghọtara n'ụzọ ziri ezi / Ọnụ ọgụgụ nke ihe odide ** Usoro ọkwa ziri ezi **: Accuracy_seq = Ọnụ ọgụgụ nke usoro ziri ezi / ọnụ ọgụgụ zuru ezu nke usoro ** Edezi Anya **: Na-atụle ọdịiche dị n'etiti usoro a buru amụma na ezigbo usoro, gụnyere ọnụ ọgụgụ kacha nta nke ntinye, nhichapụ, na arụmọrụ nnọchi. ### Njehie Analysis ** Ụdị njehie a na-ahụkarị **: - Character Confusion: Misidentification nke ihe odide ndị yiri ya - Duplicate njehie: CTCs na-emepụta oyiri odide - Length njehie: Ezighi ezi usoro ogologo amụma ** Atụmatụ Mmezi **: - Siri ike sample Ngwuputa: Lekwasị anya na ọzụzụ samples na elu njehie udu - Post-nhazi njikarịcha: Na-edozi njehie site na iji ụdị asụsụ - Integrated Approach: Ijikọta amụma site na ọtụtụ ụdị ## Nchịkọta Ọrụ ọnwụ CTC na-enye ngwá ọrụ dị ike maka ịme ihe nlereanya, karịsịa mgbe ị na-emeso nsogbu nhazi. Site na iwebata akara efu na algorithms mmemme dị ike, CTC na-aghọta mmụta usoro njedebe na njedebe ma zere usoro nhazi dị mgbagwoju anya. ** Key Takeaways **: - CTC na-edozi nsogbu nke ntinye na-adịghị mma na ogologo usoro mmepụta - Algọridim na-aga n'ihu-azụ na-enye ngụkọta nke puru omume nke ọma - Usoro nhazi kwesịrị ekwesị dị oke mkpa maka arụmọrụ ikpeazụ. - Usoro ọzụzụ na njikarịcha atụmatụ na-emetụta arụmọrụ nlereanya nke ukwuu ** Aro Ngwa **: - Họrọ usoro decoding kwesịrị ekwesị maka ọrụ a kapịrị ọnụ - Mesiri ike na data preprocessing na nkwalite usoro - Lekwasị anya na nkwụsi ike ọnụọgụ na arụmọrụ mgbakọ ● Njikarịcha nke na-adabere na ihe ọmụma nke ngalaba Ngwa nke CTC na-aga nke ọma etinyela ntọala dị mkpa maka mmepe nke mmụta miri emi n'ọhịa nke usoro usoro nlereanya, ma nyekwa nkwado dị mkpa maka ọganihu nke teknụzụ OCR.